See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/236377474

Expert system technology for the military

Article · October 1988

CITATIONS READS
22 1,900

5 authors, including:

Cora Carmody Brandon Buteau

Carmody Technology Northrop Grumman
3 PUBLICATIONS   39 CITATIONS    5 PUBLICATIONS   105 CITATIONS   

SEE PROFILE SEE PROFILE

Tod S. Levitt
George Mason University
72 PUBLICATIONS   1,596 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Cora Carmody on 10 January 2017.

The user has requested enhancement of the downloaded file.

See discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/2984108

Expert System Technology for the Military:

Selected Samples

Article in Proceedings of the IEEE · November 1988

DOI: 10.1109/5.16329 · Source: IEEE Xplore

CITATIONS READS

10 34

5 authors, including:

Cora Carmody Tod S. Levitt

Carmody Technology George Mason University
3 PUBLICATIONS 11 CITATIONS 69 PUBLICATIONS 1,203 CITATIONS

SEE PROFILE SEE PROFILE

Brandon Buteau
Northrop Grumman
5 PUBLICATIONS 61 CITATIONS

SEE PROFILE

All in-text references underlined in blue are linked to publications on ResearchGate, Available from: Tod S. Levitt
letting you access and read them immediately. Retrieved on: 02 October 2016
Expert System Technology for the Military:
Selected Samples
JUDE E. FRANKLIN, SENIOR MEMBER, IEEE, CORA LACKEY CARMODY, MEMBER, IEEE,
KARL KELLER, T O D s. LEVITT, MEMBER, IEEE, AND BRANDON L. BUTEAU, ASSOCIATE, IEEE
Invited Paper

This paper is concerned with the applications of expert systems puter technology, by the development and better under-
to complex military problems. A brief description of needs for standing of expert systems concepts, by the progress that
expert systems in the military arena is given. A short tutorial on
some of the elements of an expert system is found in Appendix 1.
has been made in sensors and control devices, and by the
An important aspect of expert systems concerns using uncertain growing need caused by an information explosion. For
information and ill-defined procedures. Many of the general tech- example, typical military command and control centers
niques of dealing with uncertainty are described in Appendix / / . handle thousands of messages in a single day and com-
These techniques include Bayesian certainty factors, Dempster- manders must decide promptly o n correct actions. Com-
Shafer theory of uncertainty, and Zadeh‘s fuzzy set theory. The
major portion of the paper addresses specific expert system exam- mercial applications i n the fields of financial investment,
ples such as resource allocation, identification of radar images, manufacturing, and business planning have similar high
maintenance and troubleshooting of electronic equipment, and information volume and time constraint problems.
the interpretation and understanding of radar images. €xtensions Although this paper concentrates on military applications
of expert systems to incorporate learning are examined in the con-
of AI, there are analogies in the commercial arena.
text of military intelligence to determine the disposition, location,
and intention of the adversary. The final application involves the While the field of expert systems i s still a long way from
use of distributed communicating cooperating expert systems for solving the military’s most persistent problems, this
battle management. Finally, the future of expert systems and their research activityfor creating intelligent machines has dem-
evolving capabilities are discussed. onstrated certain machine properties that offer great hope
and promise, particularly i n the area of manpower reduc-
I. INTRODUCTION tion. There are reasonable expectations of future com-
puters that can learn; reason; understand text and speech;
The increasing complexity of weapon systems and the
perform complex problem solving operations; recognize
growingvolume of complex information creates numerous
anomalous behaviors and warn the decision maker; under-
problems for the military [I]-[5]. Commanders must make stand drawings and photographs; and process signals such
decisions faster than ever before and maintain operational as speech, sonar, radar, and imagery.
readiness in spite of limitations o n manpower and training.
Early D O D investments [3] have helped to establish the
Artificial Intelligence (AI) technology can potentially solve
scientific foundations upon which the present U.S. capa-
many of these problems for the military, with some AI appli- bilities and thrusts i n AI and robotics are based. For exam-
cations already demonstrating their utility. The militaryand
ple, the Office of Naval Research (ONR) and the Defense
industry have made major investments in the area of AI.
Advanced Research Project Agency (DARPA) have been
This paper will concentrate o n the potential of expert sys-
supporting research in AI for over 20years through the sup-
tems, a subordinate category of AI, to the military arena.
port of “Centers of Excellence” at several prominent uni-
Within the last few years the military has witnessed an versities. These centers have published extensively, hosted
almost explosive expansion of the field of expert systems symposia for government and industry, and spawned tech-
within various agencies of the Department of Defense nological innovations such as the DARPA-sponsored Stra-
(DOD). This has been sparked by the rapid growth in com- tegic Computing Program (SCP) that has applications
including the Air Force Pilot’s Associate, the Army Auton-
Manuscript received January 6, 1987; revised May 19,1988. omous Land Vehicleand the Naval Battle Management pro-
J. E. Franklin, C. L. Carmody, and B. L. Buteau are with Planning grams. The Navy has created a Center for Applied Research
Research Corporation Government Information Systems, McLean, in AI at the Naval Research Laboratory (NRL) and has major
VA 22102, USA. programs i n maintenance and troubleshooting of complex
K. Keller i s with MITRE Corporation, McLean, VA 22102, USA.
electronic equipment, target identification from radar or
T. S. Levitt i s with Advanced Decision Systems, Mountain View,
CA 94043-1230, USA. acoustic sensors, machine learning, and fusion of data from
I E E E Log Number 8823835. multiple sensors. More recently, the Army Research Office

0018-9219/88/1000-1327$01.000 1988 IEEE

PROCEEDINGS OF THE IEEE, VOL. 76, NO. IO, OCTOBER 1988 1327
(ARO) has invested in a long-term AI research, develop- cific problem.The knowledgecan be stored in at least three
ment, and training effort with two major universities (the structures. The first i s to represent statements about the
UniversityofTexasand the Universityof Pennsylvania).The problem domain as predicate calculus (logic). The second
Army has started an AI center in the Pentagon with a con- is t o represent the knowledge as if-then rules, and the third
centration o n information management for logistics. One representation i s acollection of attribute-value pairs known
of the Army's first problems i s the correct distribution of as frames. There areother representation schemes and they
critical equipment such as cellular radios for the signal are mentioned in the Appendix. The best representation
corps. The Army has major programs i n planning and threat schemes use a combination of several of these strategies.
analysisfrom multiple sensors. The Air Force has just started Theinferenceengineaccepts the input dataand the infor-
aconsortium of universitiesoriented toward AI i n the region mation i n the knowledge base t o develop a meansof solving
around Rome Air Development Center (RADC) and has on- the problem. The inference engine may use a goal-directed
going programs i n software automation, indications and scheme that examines various potential hypotheses and
warning systems, and decision systems for military com- determines which one i s true. This technique works back-
manders. The Air Force Office of Scientific Research (OSR) wards from the goals toward the input data. Another
has sponsored research i n manufacturing science, space- approach is to start with the input data and infer conclu-
craft image understanding, systems automation through sions in a forward direction. The final approach i s t o com-
AI, and software automation. The Strategic Defense Initia- bine both techniques. The inference engine will usually
tive (SDI) i s considering the use of AI techniques t o accom- contain various search strategies that take advantage of how
plish accurate, time critical decisions and t o provide robust best t o examine alternatives and what search paths will pro-
adaptive communications for battle management appli- vide the fastest solutions.
cations. The knowledgeengineer attempts t o w o r k w i t h the actual
The remainder of this paper discusses the use of expert domain expert and represent his problem-solving tech-
systems in the military. Section I1 gives an overviewof expert niques o n the computer. The knowledge engineer is the
systems technology and what expert systems do; what the key t o the total knowledge acquisition process. Frequently,
components are; and what some of the interesting prob- the domain expert becomes the knowledge engineer, and
lems have been. Section Ill presents several examples of optimum results are achieved with fewer errors and mis-
military applications using expert systems; Section IV dis- understandings. Additional details about expert systems
cusses a learning system applied t o the military; Section V are found i n Appendix I.
considers the next step of distributed problem solving; and
Section VI provides a summary and conclusion with a
B. Uncertainty Representations and Approaches
glimpse at realistic expectations for how expert systems can
helpsolve future militaryproblemsand what research areas A popular maxim pertinent t o the study of expert systems
need t o be addressed. i s "In the knowledge lies the power." By design, expert sys-
tems augment human reasoning; it i s a natural extension
II. ON EXPERT SYSTEMS AND UNCERTAINTY
BACKGROUND t o basic reasoning approaches within expert systems t o
include the ability t o form reasonable conclusions from
A. Expert System Overview
uncertain and incomplete knowledge. Associating a rep-
Expert systems are computer programs that attempt t o resentation of our degree of belief with the knowledge con-
duplicate results obtained by actual experts i n a particular tained in a knowledge base i s one of the most common
field or domain. Recent results have demonstrated that this approaches to reasoning with uncertainty. Unfortunately,
technology can be used by the military. This section pro- most of the representations bear little resemblance to
vides a short introduction to expert systems terminology. human cognitive processes-their primaryvirtue i s thatthey
Concepts of an expert system are discussed more fully in can be implemented within an expert system and can help
Appendix I. Other sources of information o n expert sys- t o solve the problem. Appendix II contains illustrative mate-
tems can be found i n [6]-[IO]. A block diagram of a typical rial. References [Ill-[32] and the additional readings sec-
expert system i s shown i n Fig. 1. The real power of an expert tion also provide background for this research area.
system i s found in the knowledge base, which contains the The problem of reasoning with uncertainty takes o n
fundamental facts and assertions necessary t o solve a spe- greater importance when the application area concerns the
fusion of information from many knowledge sources, e.g.,
the command, control, communication and intelligence
User (C31) sensor environment. I n particular, when the inde-

r-l-- pendent knowledge sources are sensors producing low-

Knowledge Inference
Engine
level perceptual outputs, it i s essential t o have a method-
Expert Knowledge Explanation
ologyfor assigning uncertaintyvalues to evidence from dif-
ferent knowledge sources, and for combining these values
when evidence from these knowledge sources supports (or
contradicts) the same basic conclusion.
Acqulsltlon
Procedures Generally, sensor systems do not report results with 100-
c 4 percent accuracy and certainty. This i s a particular problem
when multiple sensors must be combined. For example,
when output from Sensor A support the conclusion "The
target is a type X relocatable weapon" with certainty 0.42,
and outputs from Sensor B support the same conclusion

1328 PROCEEDINGS OF THE IEEE, VOL. 76, NO. 10, OCTOBER 1988
with certainty 0.87, a mechanism should exist for combin- so that reasoning can proceed without a large amount of
ing these figures in a coherent manner. Furthermore, the initial information. The Dempster-Shafer method provides
method must deal with the situation when a similar con- a model for the gradual accrual of new knowledge into an
clusion from Sensor C i s incorporated at a later time, or expert system, and i t s inverse may be computed i n order
when even later Sensor B retracts its conclusion or updates to retract evidential contributions. There are, however, sig-
the certainty factor associated with i t s conclusion. nificant implementation difficulties i n the full scale theory
Approaches t o dealing with uncertainty often take some as discussed in Shafer [IA. The original theory calls for cre-
variation of numeric characterization. Numeric represen- ating supersets of all possible propositions and deals with
tations usually take the form of the assignment of a point independent knowledge sources. Obviously, this full
value (as the application of Bayes‘ Theorem or maximum approach will cause serious difficulties when the problem
entropy), intervals o n a range (as i n Dempster-Shafer The- area exceeds a very simple problem space, or when the
ory) or points within an evidence space. A variant o n a knowledge sources are interdependent. Ginsberg [29] and
numerical approach with foundations in set theory i s Barnett [30] have recommended ways t o solve the potential
Zadeh’s Fuzzy Logic, or Fuzzy Set Theory. A discussion of difficulties, and Lowrance and Garvey [31], [32] have suc-
these major methods is in Appendix II. cessfully implemented a slightly more advanced subset of
Dempster-Shafer theoryand applied it t o the military prob-
lem of sensor fusion in support of electronic warfare. An
C. Methods o f Dealing with Uncertainty
additional flaw in the Dempster-Shafer ruleof combination
One of the first problems with numeric methods of i s discussed by Zadeh [13], and is reviewed i n Appendix 11,
uncertainty lies in allocating initial numbers. Where do the Uncertainty Methods. The Dempster-Shafer technique i s
numbers come from?Are the numbers intended to capture used in the distributed expert system example that depicts
truth (i.e., in terms of frequency data), or are they repre- an indications and warning problem involving terrorist
senting a consensus of belief from one or more experts? activities and i s discussed in Section V.
These two interpretations lead t o different sets of prob-
lems. I n most military applications, existing frequency data Ill. APPLICATIONS
i s not available, soafoundationforfurthercalculations must
This section discusses several applications of expert sys-
be defined in another manner. There are at present no stan-
tems that are in varying stages of development for military
dard methods for laying this foundation.
applications. The choice of the applications was driven by
If the numbers are t o represent confidence or belief, the
the authors’ personal experiences and by no means rep-
initial assignment of certainties must come from the expert
resents an exhaustive discussion. The samples that were
whose assertions form the knowledge base. Different
selected represent work at the Naval Research Laboratory
experts may well assign different certainty numbers, and
(NRL), Defense Advanced Research Project Agency
may have different reasons for the assignments. The rea-
(DARPA), the U.S. Army, Planning Research Corporation
sons for assignment are left behind, once the numbers are
(PRC), Mitre, and Advanced Decision Systems (ADS).
assigned to assertions or rules. Apart from the loss of rea-
sons, there is some question as t o the quality or precision
A. Combat Management for the Marine Corps
of the initial numbers.
Another basic problem is i n the interpretation of the The Naval Research Laboratory, under the direction of
numbers. Just what does it mean t o say that “The target Slagle, developed an expert consultant system for weapon
under observation is an SS-25 missile” with certainty 0.87. allocation. The present status of the system i s a working
Does this mean that “I’m 87% sure that it’s an SS-25,” “The prototype tool. This system, called BATTLE [33], [34], gen-
probability is 87% that it’s an SS-25,“ “The odds are 87 in erates weapon allocation plans for a system with require-
100 that it’s an SS-25,” or “87 out of 100 targets that we’ve ments similar to the Marine Integrated Fire and Air Support
tracked with these kigds of characteristics turned out to be System (MIFASS).
SS-25s?” Interpretations of certainty values vary between The BAlTLE expert system evaluates the effectiveness of
degrees of belief, probabilities, betting, and frequency of individual weapons to targets and then it produces com-
occurrence. A number i s only a measure of how uncertain plete evaluation plans that consider the possible allocation
we are about a proposition and does not convey our rea- of all weapons t o all targets. Normally, this would involve
sons for doubt. exhaustive search techniques. The goal of the system i s t o
Within specific numeric methods, there are problems maximize the destruction (total value D) for all targets. I n
with single point implementations, such as Bayes’ Theo- an allocation plan, the destruction value for a target i s the
rem, since the degrees of belief and degrees of disbelief in product of the target’s strategic value and the expected per-
a proposition must always sum to one; any doubt must be centageofthetargetthatwill bedestroyed in the plan. When
represented as an unknown hypothesis. Another problem the destruction value is maximized, the plan is considered
with the Bayes approach i s that in order t o update certainty optimal. Unfortunately, achieving this optimal plan in real
values effectively, a large amount of information is needed. conditions using exhaustive search techniques can con-
Two of the applications, BATTLE and the ship identification sume too much time. For example, if we have W weapons
expert system, described in Section Ill-A and Section 111-8, and T targets, there are (T + different options. For an
use a variation and extension of the Bayesian techniques. optimal plan that considered 8 weapons and 17 targets, the
I n contrast, the Dempster-Shafer method provides a rep- computer run time, using exhaustive search, was 11 min-
resentation which separates values for belief and disbelief utes and 43 seconds. The BATTLE expert system took 6.75
(they need not always sum t o one), and a propagation algo- seconds for the same problem and achieved 98 percent of
rithm which will accept initial ignorance in a proposition, optimality. Reduction i n processing time was due t o the

FRANKLIN et al.: EXPERT SYSTEM TECHNOLOGY FOR THE MILITARY 1329

heuristic search mechanism that eliminated investigating The first step of the problem computes the effectiveness
unlikely weapon-target allocations. of a single weapon t o a single target, and the second step
A few of the 55 factors used by BATTLE t o arrive at its assessesthecompletematchupof allweaponstoall targets.
weapons allocation are shown in Table 1. BATTLE uses rules specified by a military expert to accom-
plish this matchup. A portion of the single weapon target
allocation network i s shown in Fig. 2, and this network i s
Table 1
a generalization of the inference networks of PROSPEC-
Range and position TOR. PROSPECTOR was an earlv exDert svstem that was
I ,

Personnel ieadiness developed t o assist in mineral exploration.

Counterfire ability The artillery allocation network, shown in Fig. 2, uses
Resupply
Ammunition status many different quantitative and qualitative parameters.
Number of tubes per group Quantitative parameters include reliable range estimates
Experience of enemy group for weapons or target counterfire capabilities. The quali-
Maintenance status tative parameters i l c l u d e readiness of personnel or main-
Replacement crew status
tenance status of equipment. The computational network
Physical condition
shown in Fig. 2 uses several different techniques t o com-

expert Bayesian
evidence
function
Q or

0 additional
computational
networks
Fig. 2. Weapons allocation network [33].

1330 PROCEEDINGS OF THE IEEE, VOL. 76, NO. IO, OCTOBER 1988
bine information prior to allocating weapons to targets. i s allowed and a weapon can be fired only one time. The
Classic Boolean operators of and and or are used to com- weapon and target assignments are shown on the left-hand
bine information such as resupply alternatives and coun- side, and the destruction value i s shown on the right-hand
terfire capability. The Bayesian evidence functions indi- side. Fig. 3 illustrates exhaustivesearch with an optimal allo-
cated by the triangular symbols are used to combine cationof W4toT2, W3toT3, W2toT1,and WltoT1,yield-
information such as friendly versus target readiness, ing a total destruction value of 401. Fig. 4 shows the same
friendly and target supply status, or target appropriateness.
In this case, there i s a degree of uncertainty in the infor- F I G W OF
CHOICE ~OICE CHOICE CHOICE MERlTv-
mation that i s used as well as uncertainty in how this infor-
(1)-W4->T1-(2) -I -W3 >T3-(3) -I W1 ->T2--Wz->T1 - W 2 - > T z 389
mation should be combined. As discussed earlier, and in I 387
Appendix II, the Bayesian technique provides a means to I I !W l - > T l
i I
combine thesevarious forms of uncertainty. Expert defined I !w 1 - > n
heuristics are indicated bythefilled-in symbols.These heu- I
ristics usually represent empirical results that aid in reliable -i W 3 - > T I
weapon to target allocations.
A typical scenario for the allocation of 4 weapons to 3 tar-
gets is shown in Table 2. In this example, the target’s stra-

-w4->n
Table 2 I
I
I
Target TI T2 T3 I
I
Strategic value 200 100 150 I
Weapon effectiveness
w1 60% 60 % 60 %
w2 70 % 30 % 0
w3 80 % 0 90 % i
w4 90 % 90 % 90% I
I
I
- W4 - > T2 -W3- I
> T1

tegic value i s shown along the top, and the weapon effec- I
I
tiveness is the expected percentage of the target that will I
I

be destroyed by that weapon. The allocation tree i s shown

in Fig. 3 for the case in which every weapon i s used in every
allocation, massing of multiple weapons on a single target
W2->T3
mo
uRE OF
mvAUIB Fig. 4. Pruned allocation tree using heuristic search.
389
387
357 tree but indicates that the total number of computations
333
was reduced dramatically by the addition of afew heuristic
354
338 rules to the search tree. This pruning was accomplished by
316 some very simple considerations. One consideration was
289
to evaluate the best weapon to target allocation to that point
268
259 and then to use some look-ahead heuristics to determine
228 if the next branch of the tree under evaluation could pos-
200 sibly exceed the existing best weapon-to-target destruction
383
367 value. If the estimated cumulative destruction value did not
349 exceed the existing value, this branch of the tree was no
330 longer evaluated.
334
332
Frequently, there are situations in which all of the infor-
349
mation i s not available for the expert system to perform the
325
289
required allocation. For this case, the expert system must
ask questions of the user. The BATTLE system uses a new
299
221
179
technique, merit, to ask the highest payoff questions first,
and to minimize the user’s timeand cost to provideanswers.
268
343 The merit system performs this function by considering how
285 easy it i s for the user to provide the information and how
277
useful the answer will be for the weapon-to-target alloca-
284
257 tion.
w371
The merit system starts with the concept of the self merit.
The self merit i s defined as the expected variability in the
348 answer to a question divided by the cost to acquire the
231
232 answer. The potential variability in the answer to aquestion
Fig. 3. Example of weapon allocation to targets and sub- would be very high, for example, with matters associated
sequent figure of merit value. with weather since it i s expected to change frequently. The

FRANKLIN et al.: EXPERT SYSTEM TECHNOLOGY FOR THE MILITARY 1331

cost is associatedwith the difficulty in providing the answer. 25 questions to achieve this same level of performance as
One measure of cost could be the time it takes t o find an the merit strategy. The expert system must minimize the
answer to a question. Therefore, the self merit of a node questions asked of a Marine artillery officer when he i s i n
involving weather would be relatively high because of the a stressful combat environment, and the merit strategy
large expected variability of the response divided by the helps reduce unnecessary questions.
relatively low cost of acquiring weather information. The description of this project has been used as the initial
Fig. 5 illustrates the merit system. The self merit asso- exampleto showtheapplication of AI t o a real militaryprob-
ciated with the node C1 i s the absolute value of the change lem, in this case one of resource allocation. Several points
should be mentioned. First, the implementation of expert
systems i s at best an empirical problem. There i s n o sure
tried and true technique that i s universally applicable. In
this case the PROSPECTOR paradigm was useful, but it
needed additions and alternatives such as the merit ques-
tioning strategy and a pruning technique that was custom-
ized t o the problem. This project demonstrated that the
resource allocation problem was one that could be useful
for the military. The next project t o be discussed used many
I"di:] of the tools developed by BATTLE and applied them t o a
ship identification effort. I n fact, one of the things being
Fig. 5. The merit concept. studied at the time was the question of whether the BATTLE
expert system could be used in other domains. This i s avery
real problem in the extensibility of expert systems. Fre-
in the response parameter (AP,), divided by the cost (AC,).
quently they can work fine in one problem domain but are
The next important consideration i s the effect of this answer
at node G1 and the impact it will have o n the next higher utterly useless in an apparently similar problem domain.
node G. This i s referred t o as the edge merit and it i s illus-
trated in Fig. 5. It i s theabsolutevalue of thechange in node B. Applications o f AI for Ship Identification
G(AP) divided by thechangeat node Gl(AP,).Thetotal merit
Another project at the Navy Center for Applied Research
at node Gassociated with G1 i s the product of the self merit
in Artificial Intelligence concerns the correct identification
and the edge merit as shown in Fig. 5.
of ships from images. The methodologies used in this expert
Fig. 6 shows acomparison of the merit strategyfor asking
system are spin-offsof the BATTLE system discussed above.
A prototype decision aid has been developed that dem-
0.25 onstrates the feasibility of using AI techniques t o help
human interpreters recognize images of ships. The heu-
ristic reasoning processes used by an expert analyst were
0.20
modeled t o perform this task. The prototype system uses
heuristic rules provided by a human expert to determine
L 0.15 the most likely classification, given the available feature
g
0
v)
data. The system has been tested o n over 100 images of 10
z 0.10 similar classes of ships, and it provides the correct iden-
tification 84 percent of the time [35].
0.05 Accurate identification of ships in real time can be a very
difficult task. It i s desirable t o identify these ships within
minutes of detection. The images are hard t o interpret for
numerous reasons such as weather, turbulence, lighting,
3 6 9 12 15 18 21 24 27 interference, and viewing angle. When thesedifficulties are
Number of questions
coupled with the hundreds of classes of ships that sail the
oceans and the time constraints to make accurate deci-
Fig. 6 . Comparison of questioning strategies.
sions, we can understand the need t o help operators as
much as possible. Experience has shown that even expert
questions with two other techniques. The other two strat- operatorswill have bad daysor beso burdened with numer-
egies are the random selection of questions and the fixed ous target detections that they will become less accurate
order strategy. The fixed order strategy i s calculated prior and on occasioneven missobviousclues.Theaverageoper-
to the availability of answers to any questions and it is not ator is frequently inconsistent and uncertain. He also has
updated as answers become available. This reduces the a difficult time determining what should be the next clue
need for online computation but it results in a fixed order to be pursued to reduce the target class ambiguity and
in the set of questions and frequently takes more time. The finally to narrow his decisions t o a specific selection. The
erroron the Yaxis refers to thedifference between theopti- thrust of the NRL research headed by Booker [35] was to
mal plan that maximizes the destruction valueand the actual allow for the use of automatic equipment that performs sig-
destruction valueforthethreestrategies.The merit strategy nal processing and pattern recognition and combine this
arrived at a solution that iswithin five percent of an optimal equipment with AI techniques to aid the operator. It was
plan and required less than a dozen questions. The other very important to have the expert system capture the oper-
two strategies used in the comparison would require 18 to ator's judgment, pattern recognition capabilities, the

1332 PROCEEDINGS OF THE IEEE, VOL. 76, NO. 10, OCTOBER 1988
 e
lr“““..
Explanation
Executive

I I
Network I Interpreter
I

VERTICAL - NO TAPER
LINEARTAPERFORWARD
OFFORWARD MAST

-I
Fig. 7. Expert system design.
Inference Network
I rn SLIGKTNONLINEAR TAPER

knowledge and the reasoning that he uses for his expert

analysis.
The block diagram of the system is shown in Fig. 7. The
system guides the operator t o investigate specific clues or
parameters and then suggests an identification for a spe-
cific unknown ship. The Explanation System uses the logic Fig. 8. Mast taper descriptions.
provided bytheexpert and can showtheuserwhytheexpert
system made a specific decision regarding the identifica-
tion of an image as a specific ship class. The Questioning 9. This network i s based o n the techniques used in PROS-
System decides what question should be asked next. It uses PECTOR [37] and in BATTLE as discussed earlier. Fig. 9 i s just
the merit strategy, which is also used i n BATTLE, tooptimize a small subset of the complete computation network. Here,
theorder of thequestions.TheVolunteer System allowsthe as in BATTLE, the network combines Boolean operators and
operator to provide as much initial data as he deems fit. It Bayesian evidence functions. The inputs t o this network are
also allows the operator t o volunteer additional informa- related to thevariouscluesthat the imageexpertfound use-
tion during the questioning period. The Propagator cal- ful. As can be seen from the figure, these clues are asso-
culatestheeffectofthe uncertaintyofthedataandtheoper- ciated with locations of masts, locations of superstructure
ator’s observations as the initial information travels blocks, and the presence of linear and nonlinear tapers.
throughout the computational network toward the goal of The order of the questions i s changed dynamically by the
ship identification. The techniques used here are a varia- merit strategy developed at NRL and used for the first time
tion of the Bayesian techniques described earlier and in in BATTLE. This technique that was described previously
Appendix II. The Inference Network combines the initial has been shown t o reduce the required number of ques-
observations and clues into intermediate hypotheses that tions needed t o identify a ship. Excerpts of an actual session
are combined further towards a final conclusion. Fig. 8 i s using this expert system are shown below. The expert sys-
a simple cartoon that the image expert found very useful tem’s questions and comments are shown in bold font.
in explaining his reasoning t o the AI knowledge engineer.
Concepts such as, “do the masts have a linear taper,” as Expert system question 1: To what degree do you believe
shown in the top of Fig. 8, were important to the image that the profile-has-isolated-directors-fwd-of-bridge-with-a-
expert’s approach to identifying specific ships. Fig. 8 depicts linear-taper?
other concepts such as nonlinear tapers, linear tapers with
Human response t o question 1: ( - 3)
a gap, and no tapers. For example, the system might ask if
Note that the operator does n o t believe that this condition
there is a linear or nonlineartaper in the forward mast struc-
i s present.
ture, and the answer would be an input t o the expert sys-
Expert system comment: I neither suspect nor doubt the
tem. After the operator is asked about the presence of a
possibility of ForestSherman, Bainbridge, Sverdlov, Coontz,
specific clue, such as ”is there a linear taper in the forward
Leahy, Belknap, Truxtun, California or Virginia.
section?”theoperatorwilI respondwith anumber between
I doubt that image84 could be LongBeach
- 5 and + 5 . A - 5 indicates that the operator is sure that
this characteristic does not exist, and a + 5 indicates that Noteat this point thatthe machinealreadydoubts the Long
the operator i s sure that this clue i s present. A zero would Beach ship class i s likely and i s neutral about the rest.
indicate that the operator does not know if that feature i s
present or not. I n this way the operator can specify his Expert system question 2: To what degree do you believe
uncertainty. These answers are then used as inputs t o the that the profile-has-a-gap-between-aft-mast-and-aft-super-
computation network. A typical network i s shown in Fig. structure-block?

FRANKLIN er al.: EXPERT SYSTEM TECHNOLOGY FOR THE MILITARY 1333

 OTH MASTS FIT VA

EDGE OF SUPPORTING
N SUPPORTING BLOCK ONLINEAR TAPER FWD
I I I

Fig. 9. Typical inference network.

Human response t o question 2: ( + 5) total images that were from the chosen ship classes that
Note that the operator strongly believes that this clue is were represented i n the expert system. The other 18 were
present. for ships not contained i n the chosen category and the
Expert system comment: I strongly suspect that image84 is expert system did not know of their existence. The success
Coontz rate was 84 percent. I n 85 of the 101 trials, the ship class
I doubt that image84 could be ForestSherman, Sverdlov, ranked first by the expert system was the correct identifi-
Leahy, Belknap, Truxtun, California or Virginia cation. I n 94 of the 101 trials, the correct target class was
I strongly doubt that image84 is Bainbridge or LongBeach listed as a plausible identification. The expert system pro-
vided consistently useful discrimination between the top
Note at this point the machine has asked a question using
ranked class and the second rated class. The summary of
the merit strategy t o determine the most useful question
this discrimination i s shown below.
t o answer. Indeed, with the response from the user, it i s
already focusing o n the Coontz class. Correct trials:
Average confidence i n top ranked class was 1.7.
These questions continue, and at the end the machine Average confidence in second ranked class was -2.4.
responds with its list of potential classifications (identifi- This means that the system usually comes u p with only
cations) i n a priority order. I n addition, the machine gives one plausible classification.
its confidence of these classifications as shown below. incorrect trials:
Average confidence i n top ranked class was -0.16.
Class Confidence Class Confidence Average confidence in second ranked class was -0.99.
Coontz 4.9 Virginia -5.0 Note that i n 84 percent of the cases where the machine
California -4.8 Sverdlov -5.0 correctly identified the ship, the expert system average con-
Bainbridge -5.0 Leahy -5.0
fidence was 1.7 and this indicates a relatively high positive
Truxtun -5.0 ForestSherman -5.0
Belknap -5.0 LongBeach -5.0 belief that the chosen t o p ranked identification i s correct.
Further, the expert system had a confidence of -2.4 in the
second ranked class and this indicates a fairly strong belief
Expert system comment: I strongly suspect that image84 is that this i s not the correct class. This helps the user believe
Coontz in the expert system’s results because of the large sepa-
I strongly doubt that the image84 is ForestSherman, Bain- ration of +1.7 and -2.4. I n the cases where the machine
bridge, Sverdlov, LongBeach, Leahy, Belknap, Truxtun, Cali- incorrectly identified the image, i t s confidence i n the first
fornia or Virginia choice was -0.16. Clearly, the machine was not “sure” of
its choice, and the user can see this and would probably
Note that the expert system has listed the Coontz first
want t o acquire a new image and add more results prior t o
with a confidence factor of positive 4.9. The California class
a final selection. More detailed results can be found in a
was listed second, but the expert system was not very con-
publication by NRL’s Booker [351.
fident of this because it has a confidence factor of -4.8.
Clearly it did not believe it was the California class or any
C. Maintenance a n d Troubleshooting
of the others that are o n the list. This i s an extremely useful
feature since it can help the operator decide if he i s satisfied Since the early 1960s military equipment has steadily
with the expert system‘s conclusions. I n this case, theexpert increased in complexity and variety, while at the same time
system was very confident that it was the Coontz and none the pool of trained technicians has been decreasing. A major
of the others. cost of operations i s i n fault diagnosis and repair, the pro-
There were 119 images investigated. This included 101 curement of maintenance equipment, and the training of

1334 PROCEEDINGS OF THE IEEE, VOL. 76, NO. 10, OCTOBER 1988
technicians and operators. Each of the armed services has
problems which are unique t o i t s mission, but all share Prior model of unit
Generator
problems of space, difficulty in providing logistics support, under test
I I manuals/experts)
and limited technical manpower. These factors coupled
with the demands of operations place heavy emphasis o n
speedy and accurate diagnosis and repair in the field. These
difficulties have created prime opportunities for the appli-
cation of AI, and a number of efforts are underway.
All three services are investigating alternatives using AI
in the area of Maintaining and troubleshooting electronic
equipment. The Air Force has two major programs. The first
I
Knowledge
Initial rule base
(expert)

Base
is the Integrated Maintenance Information System (IMIS)
project that i s designed t o give flightline personnel access
to all onboard diagnostic data as well as maintenance and Fig. 10. Knowledge base compiler
scheduling records. The second program i s the Generic
Integrated Maintenance Diagnostics (GIMADS) effort that come from the initial rule base. Some sources of this infor-
proposes to use AI and conventional techniques to address mation are block diagrams of the actual equipment indi-
thediagnostics problem for an integrated system. The Navy cating the interconnection and the functionality of the sys-
has two major programs, Integrated Automatic Test Gen- tem components. The basic information about the circuit
eration (IATG) and the Integrated Diagnostic Support Sys- can be combined with generalized circuit analysis tech-
tem (IDSS). The IDSS will use expert systems t o assist the niquesandfaultdiagnosistechniques. In thisway, new rules
equipment technician t o troubleshoot by providing opti- can be generated by the expert system from an under-
mal test trees that are adaptive to the changing conditions. standing of the circuit and the problem. Reliability and
The following examples of AI being applied t o the field of maintainability analysis should also be used. This would
maintenance and troubleshooting will draw heavily on the include statistical data associated with past failures as well
efforts of DeJong, Pipitone, Shumaker, and Cantone [36], as failure prediction algorithms.
[38]-[41] at the Navy Center for Applied Research in Arti- One of the new features of the approach by Pipitone [411
ficial Intelligence. Other relevantwork includes Duda etal. isthat he has included relationshipsof h o w t h e module will
[37l, DeKleer [42], and Davis et a/. [43]. react functionallytovarious faults. An example for avoltage
There are some basic technical issues that should be controlled oscillator (VCO) would be that an input voltage
addressed when AI i s applied t o maintenance and trou- that is outside of the desired range will cause the output
bleshooting. It i s not a straightforward transition from an to be outside of the desired frequency range, in the same
application of medical diagnostic successes such as MYClN direction and with the same relative magnitude. Another
to electronic troubleshooting, because too many unex- causal relation could specify that if the output load i s
pected causes of equipment malfunctions cannot be antic- reduced below a certain value the signal will decrease.
ipated in atraditional if-then rule-based backward-chaining Another example associated with a high pass filter would
paradigm. For example, solder splashes that short out sev- be that if the output of a high pass filter is a high dc voltage,
eral components o n a printed circuit board, a cable that is the high pass filter i s very likely t o have failed. More exam-
crushed because someone rolled a heavy piece of equip- ples can be found i n recent publications [38]-[43]. One of
ment onto it, or a piece of equipment that was doused with the desirable features of an approach using a causal func-
a pot of hot coffee are typical unfortunate circumstances tional description of each module i s that it can be used with
that sometimes approach the routine. All of these condi- slight modifications when describing a similar module in
tions cannot be anticipated with appropriate rules, and thus another system. Recent results indicate that the addition of
a more general methodology must be used. this causal functional knowledge requires fewer tests than
As an insight t o the reader, the following description of an approach that uses n o functional information.
an actual expert system at NRL will be used. The system Fig. 11 shows one configuration for the block diagram for
design was evolutionary and several distinct systems were the diagnostic system. The knowledge base shown at the
developed and tested. Fig. 10 shows how the knowledge top i s the same one that was discussed and shown in Fig.
base was generated i n the first NRL system. The initial rule I O . The inference engine investigates and chooses what
base, as shown in this figure, consists of the heuristics, ad rules should be used from the rule generator in Fig. IO. If
hoc rules, and information about the specific design and there is n o appropriate rule, it will generate one that can
the operating environment. I n this way the knowledge of be used t o assist in the troubleshooting. An important con-
the designer, the operator, and the technician can all be sideration for this expert system i s t o decide what i s the next
tapped, and the results of their experience and understand- best test to be performed.
ing of the problem can be exploited and put into the initial The next best test should consider what test will cost the
knowledge base. This initial rule base will not be adequate least and will result in the most information gain. Initial
t o resolve all of the troubleshooting problems that will be attempts at calculating the next test used game theory
encountered i n the field. As an augmentation t o the system, derived from earlier AI research.Thistechniquewould con-
a more general source of information will be used, which sider the machine as the opponent, similar t o an opponent
i s shown i n the upper right-hand corner of Fig. I O . This por- in a chess match, and the strategy would be to maximize
tion of the knowledge base must use other sources of infor- your gains and minimize the anticipated losses caused by
mation that can be used t o generate rules by the expert sys- your opponent’s move. This proved to be corn putationally
tem. These rules should be i n the same form as those that infeasible [36]. Next, a Gamma Miniaverage [38] technique

FRANKLIN et al.: EXPERT SYSTEM TECHNOLOGY FOR THE MILITARY 1335

 1
how evidence extracted from SAR imagery and terrain data-
bases i s matched against the models and combined t o infer
I By
Knowledge
I the presence or absence of military forces. In particular,
radar data, forces, terrain or other entities that have been
modeled i n the certainty calculus can be used as evidence
for inferences output by the system. This i s the basis for a
clean split between inference and control in ADRIES.
The model databaseof ADRlES implicitlycontainsall pos-
sible chains of inference that the system can use t o draw
any conclusion. However, any information whatsoever can
to testhepair be used for control i n the system. As an extreme example,
Test if a human intelligence report was available that indicated
Technician
the presence of a force in a region, ADRIES could use that
to direct processing of imagery and terrain of that region
without danger of circular reasoning in itsconclusions. This
Fig. 11. Diagnostic system.
is becauseany output hypotheses about forces in the region
must be supported by image and terrain evidence as spec-
ified in the model database, and that inference is neces-
by Slagle was used. I n this case, average losses were cal-
sarily from the certainty calculus.
culated, since the equipment was not actively plotted
As another example, other source (other than SAR
against the technician (although some technicians have
imagery) intelligence data can be used t o direct search,
argued t o the contrary). The miniaverage approach was still
make predictions, or act as a trigger t o activate agents or
computationally too costly. Pipitone [361, [391-[41] intro-
other processing, but it cannot be fused t o provide evi-
duced heuristic screening at first and then used a one level
dential support for system outputs unless it i s modeled in
miniaverage computation. The results in the Pipitone work
the model database and in the certainty calculus. It follows
have shown that functional knowledge is a necessary com-
that to extend ADRIES t o a full fusion system, it will be nec-
ponent of the troubleshooting process and that the total
essary t o do the research o n the knowledge representation
number of necessary tests can be reduced. I n addition, the
of other source information, and the probabilistic rela-
functional knowledge can be used t o show the technician
tionships between other source intelligence and the en-
how the test can be done and what results should be antic-
tities already accounted for in models and the calculus.
ipated.
Fig. 12 shows the concept behind the model-based Baye-
The progress to date has shown that expert systems can
help in the area of diagnostics and troubleshooting. The
first expert systems will be used as decision support sys- MQ!XL !MAG€
tems t o aid the technician. As progress i s made, the state
of the art will allow the use of expert systems t o be used
i n an autonomous mode i n applications such as generating TANK COMPANY
automatic test equipment code or space-based mainte-
nance and troubleshooting.
TANK VEHICLE

D. Advanced Digital Radar Imagery Exploitation System

(ADRIES)
The Advanced Digital Radar Imagery Exploitation System
(ADRIES) is a software prototype testbed for research o n
extraction of information from radar imagery. Its objectives EAYES NFT

are t o provide a system for enhancing and automating var-

--Y
BATTALION NOT BATTALION
ious aspects of digital radar image interpretation, for appli-
cations including tactical intelligence missions, military sit-
uation assessment, and target recognition. ADRIES is COMPANY NOT COMPANY

capable of producing interpretations of the possible mili- Fig. 12. ADRIES inference methodology.
tary situations with a set of radar imagery, collection param-
eters, a priori terrain data, such as maps or digital terrain
databases, and othertactical data.ADRIES iscurrently under sian inference i n ADRIES. Evidence, such as clusters of
development. Keyworkon terrain, detection, and non-radar detections, are matched against the geometry of forma-
intelligence source reasoning will be presented in future tions that are explicitly represented i n the model database.
publications. Here we emphasize intermediate results i n Matches lead t o generation of alternative hypotheses of the
knowledge and model-based military unit inference. presence of forces, such as batallions and companies, that
ADRlES i s founded o n a theoryof model-based, Bayesian are dynamically created (as imagery is processed, for exam-
probabilistic inference. Models represent knowledge of the ple) and instantiated in a data structure called a Bayesian
organization and formations of military units, and they also network. Terrain rules are applied to the terrain extracted
specify how knowledgeof terrain provides evidence in sup- from terrain databases of the area corresponding to the
port or denial of the presence of types of forces at given imagery, and the certaintycalculus combines that evidence
locations. A probabilistic certainty calculus [44] specifies with the model match in computing the probabilities asso-

1336 PROCEEDINGS OF THE IEEE, VOL. 76, NO. IO, OCTOBER 1988
ciated t o the hypothesized forces in the Bayesian network. cessing plans from the Agenda agent, and performs
In practice it does not matter in what order the terrain evi- resource allocation to distribute processing o n the distrib-
dence probabilities or goodness of formation match prob- uted system. It sends back a summary of its process mes-
abilities are calculated, as the certainty calculus will obtain sages t o the Control/lnference agent.
the same values regardless of the processing order. The cal- Imagery Location Agent: The Imagery Location agent i s a
culations used for computing prior and runtime probabil- spatially indexed database of imagery in ground coordi-
ities in the calculus are presented in [45], [46], [44]. See also nates. It keeps track of virtual sub-images cut from larger
[47l. images, and maintains message-level records of the pro-
I n the course of the ADRIES program, theoretical work cessing done to imagery. The imagery files are not resident
has also been performed on detection algorithms [48], [49], with this agent; in particular, the Imagery Location agent
clustering algorithms, and elicitation of terrain rules and does not send or receive actual imagery.
probabilities [45], [46]. All of this work is incorporated in Registration Agent: It performs coarse registration t o
functionality of the relevant distributed processing agents. compute the ground coordinates of an image based o n its
ADRIES is built as a distributed set of software agents platform parameters and flat world assumptions. The Reg-
communicating by message passing. The agent decom- istration agent also computes a refined registration of the
position for ADRIES i s pictured in Fig. 13. There i s also a set image t o the ground. For a given image, it computes a func-
of databases used by multiple agents in their processing. tion that takes the elevation at a point in the terrain and
These are pictured in ovals in Fig. 13. All agents have access, outputs the corresponding point in the image.
either directly or indirectly, to all databases. Individual Lo-Res Detection Agent: This agent detects potential
agents may have additional databases. In the following, we vehicles in low resolution imagery. It also computes the
briefly summarize the functionality of each agent. likelihoods corresponding t o the hypotheses vehicle ver-
Controlllnference Agent: The Control/lnference agent sus non-vehicle.
plans system processing t o fulfill the exploitation requests Clustering Agent: The Clustering agent takes detections
(ER) received from the user interface. Basically, an ER spec- and their probabilities as inputs, and outputs clusters of
ifies what forces t o look for in which geographic locations. detections and the probability that the cluster contains a
It posts its processing as messages sent t o the Agendaagent. military unit of "array" size (e.g., 8-15 vehicles). I t accounts
The Control/lnference agent maintains the Bayesian net- for inter-vehicle spacings, likelihood of false alarm detec-
work and decides when to generate a new hypothesis in the tions, and dispersion of the cluster o n the ground versus
Bayesian network based o n the available evidence. It also the expected extent of the array-sized formation.
decides when to terminate processing o n an ER. Spot Mode Detection Agent: This agent performs vehicle
Agenda Agent: The Agenda agent receives process plan detection on high resolution imagery. It also computes
messages from the Control/lnference agent and sends them probabilities of vehicle versus non-vehicle.
o n t o the Exec agent. It provides a loose coupling between Vehicle Classification Agent: This agent performs vehicle
the planning i n the Control/lnference agent and the recognition on high resolution imagery, and alsocomputes
resource allocation in the Exec agent. probabilities over the set of hypotheses of the vehicle type.
Exec Agent: The Exec agent picks u p the current pro- Other Source Agent: This agent provides signals intelli-

1 I
SPOT MODE vwcLE OTHER
CLUSTERM DETECTlON CLASSIFICATION SOURCE

7
(2
JDATABASES

Fig. 13. Agent decomposition.

FRANKLIN et al.: EXPERT SYSTEM TECHNOLOGY FOR THE MILITARY 1337

gence and other source intelligence summaries indexed by that is, they useas pre-conditionsterrainattributesthatcan-
geographic location and force type. not be gotten by direct access from the terrain database.
Focus Agent: Focus takes as input a geographic region No probability model currently exists for these rules.
and aforce type and outputs a prioritized list of sub-regions Instead, they interact with formation matching to adjust
for search for forces of that type in the region. Focus uses expected force formations based on terrain and military
inputs from the Terrain agents and the Other Source agent. constraints.
Matching/Prediction Agent: Match ing/Prediction takes as User Interface Agent: User Interface allows the user to
input the locations of hypothesized forces and matches interactively input an exploitation request. It also displays
them against doctrinal formations. It also interacts with the all system results, including imagery, terrain overlays, and
Terrain agent to adjust military formation models to fit the multi-level force hypotheses. The user can interact with the
underlying terrain according t o military doctrine of deploy- outputs to obtain explanations for the system’s conclu-
ment. Predictions are provided by reasoning about the sions. .
forces missing from partial matches. Matching/Prediction Inference in ADRIES i s performed over a space of hier-
also provides the probabilities for goodness of fit t o for- archically linked hypotheses. The hypotheses typically
mation used as evidence by the Control/lnference agent. (although not solely) represent statements of the form
Terrain Analysis Agent: Terrain Analysis i s currently per- “There i s a military force of type F in deployment D at world
formed over three different terrain rule bases. Local Terrain location L at time T.”The hierarchy in the hypothesis space
Analysis takes as input a force type and ageographic region corresponds to the hierarchy inherent in military doctrine
and uses terrain rules t o compute sub-regions of constant of force structuring. Thus, ”array-level” hypotheses of mil-
probability indicatingthe likelihood thataforceof that type itary units such as companies, artillery batteries, and mis-
will be located i n the sub-region. The terrain rules used are sile sites are linked to their component unit hypotheses of
those that take as pre-conditions terrain attributes that can vehicles, artillery pieces, and missile launchers. Similarly,
be gotten from direct access to the terrain database. The moving upward in the force hierarchy, companies are
reason for separating these rules from ”structural” terrain grouped to form battalion hypotheses, battalions to form
rules i s that we have created a probability model for local regiments, etc.
terrain rules. Thus, these probabilities can be combined as The hypotheses are generated by hierarchical and partial
evidence bytheControlllnferenceagent as part ofthe prob- matching of military force models to evidence available in
ability of output hypotheses. radar imagery. Thus, ADRIES is a model-based radar vision
The Mobility Corridors sub-agent takes as input a force system. Evidence of the truth (or denial) of a hypothesis i s
type and geographic region and outputs sub-regions accrued numerically from probabilistic estimates about the
through which a force of a given type can move across the sub-hypotheses that comprise their parent hypothesis.
entire region.This sub-agent i s intended to be upward com- Although control of processing in ADRIES can be com-
patible with a future Avenues of Approach agent that does plex, the structure of inference follows a pattern based on
more global reasoning over the tactical situation. the models of military forces. These models consist of force
The Structural Terrain Analysis sub-agent takes as input types (i.e., names) and spatial-geometric deployment data
a force type and a geographic region and uses terrain rules relating force types. The part-of relationships for generic
to compute sub-regions that are acceptable for a force of force models are shown in Fig. 14. Numbers next to forces
thegiven typetooccupy.These rulesare relational in nature; indicate how many of a given force are expected as com-

-+-
MOTOR PARK SIGNALS GROUP w E w n w s

HEAWUARTERS (1)

COMPANY (3) HEAWUARTERS (1)

t
VEHICLES 12-66)
VEHICLES (9-12) HEAWUARTERS

t
VEHICLES (1)

Fig. 14. Part-of hierarchy for generic forces.

1338 PROCEEDINGS OF THE IEEE, VOL. 76, NO. IO, OCTOBER 1988
Fig. 15. F o r t c deployiiic3ntiniodcli. ( a i D e t c n i i v e Ihittalion tormation. ( b )D(1tcXnsive c o n i
panv iorniatioii. ( ( i Biitt,ilion c o n s t r a i n t \ . id) Rry,inieiit c o n \ t r i i i n t s .

ponents ot the parent torce. Fig. IS(& shows a standard straints tor a battalion and regiment. Fig. 16 illustrates the
detensive deployment tormation lor a battalion, while Fig. internal representation 01 the constraints in the force-
15(b)shows the tormation associated to the pattern ot w h i - deployment models as semantic nets. At the top of each
cles making up a company. Models are also represented constraint is a parent model item that is represented by a
along "is-a" or type relationships by appropriately weak- rectangular box. A box with rounded corners represents
ening model constraints. For example, a "tank company in the relation and its associated parameters. Both parent and
detensive deployment" is-a "tank company" which is-a subordinate model items can be forces or formations.
"company,' which is-a "array-level torce." ADRIES has a utilityand model-based approach to control
The models are represented as constraints in semantic ot force-deployment pattern matching. Algorithms are
nets. A constraint consists o i a relation name, parameters employed opportunistically, as determined by the utility01
associated with that relation, and additional model attri- tasks in light of thecurrentlyavailabledata, ratherthan nec-
butes. Fig. 15(c)and (d)give pictorial representationsot con- essarily being checked or applied in a fixed order. Some

\ /

Fig. 16. B a t t a l i o n $ernantic net

1339
example rules, implicit in the model-base, and inference evidence to a system output, along with an associated prob-
chains are pictured in Fig. 17. (The inferencechain pictured ability that supports that result.
uses more models and rules than are shown in the figure.) We selected probability theory as the underlying tech-
The utilitytheory i s derived from decision analytic methods nology for this numerical accrual of evidence. One of the
on topof the Bayesian inference. For detailsof the approach, major motivations forthis choice i s that Bayesian inference,
see [57, [58]. a well-developed scientific theory, already exists for prob-
abilistic evidential reasoning; see, for example [50]. This
approach requires us to lay out, apriori, the links between
evidence and hypotheses in the models over which the sys-
IF (SA-6) THEN (search for nearby HQ) tem will reason. Having laid out these links, we then need
IF (HQ) THEN (generate hypothesis Regiment or Division) a numerical interpretation of the conditional belief, i.e.,
IF (partially matched higher-level-force hypothesis)
probability, in a hypothesis, given chains of evidence that
THEN (use formation-deployment models to predict
support it through links. This is similar in spirit t o propa-
unobserved sub~lorces) gation networks [51], t o influence diagrams [52], and other
probabilistic accrual models. Hierarchical Bayesian infer-
Inference/trim ence was introduced by Gettys and Willke [53], Schum and
SA-6
Ducharme [54], Kelly and Barclay [55] and has been carried
/
Search for HQ forward by others, e.g., [44], [56], [57. Fig. 18 shows the

SYMBOLIC MILITARY UNIT INFERENCE

REGIMENT

Cr,t,calcues [ Use formation models

and predict unobserved sub-forces BATTALION

ARRAY

I It$
M a t c h L e d forces in data
VEHICLE
Good match
A
Accrue certainly
No match

Adjust formation
model for local
I U I

of hypotheses
terrain
INFERRED
Fig. 17. DRIES knowledge-based processing. ......................... CLASSIFICATION

The concept of conflict resolution and numerical accrual

of evidence is that, while it i s often the case that automated ..................... VEHICLE DETECTION 9
PROCESS
matching can be locally incorrect, the weight of the global
evidence will override local mismatches and result in an Fig. 18. Hierarchical hypothesis-space relationships.
unambiguous interpretation of the tactical situation. The
global picture is presented in ADRIES as a set of goals about
what it expects to see in the SAR imagery. For example, a hypothesis and evidential accrual hierarchy for Bayesian
command to verify the presence of a division sets up a goal inference in ADRIES.
to find a division. The Controlllnference agent then looks The evidential accrual for the generated hypothesis i s
up the structures according to the local terrain and military performed by using a Bayesian network. This paradigm sup-
situation. The model i s used t o set up sub-goals recursively ports the propagation of evidential probabilities through
to find the component military units of the division. The a hierarchical tree of hypotheses in a coherent, stable, and
existing force-deployment hypotheses are stored as nodes efficient manner [51]. The result of the evidential accrual i s
in a Bayes net. These represent the goals and subgoals that a posterior belief value for each hypothesis indicating i t s
have already been pursued or have been satisfied. relative likelihood. Mutually exclusive conflicting
Thus, the goal structures and the force-deployment hypotheses are collected in a single Bayes net node. Each
hypothesis space are dual t o each other. In the course of node then consists of a belief vector indicating the current
processing, control moves back and forth between them. relative belief that each of theconflicting hypotheses i s true.
Goal structures help predict where t o look for forces in the For example, a node can group the beliefs for acluster being
imagery and the force deployment hypothesis space. Hav- a battalion or being a false alarm (i.e., not a battalion). The
ing localized a search area, data is retrieved from the Bayes links connecting the nodes consist of conditional proba-
net and matched against the models associated t o the goals. bility matrices indicatingthe relationship of the hypotheses.
Matches maytrigger bottom upactions to infer other forces These probabilities are derived from terrain constraints and
hypotheses, which may in turn trigger the generation of locality of the hypothesized forces. For example, for a Bayes
new goals. link connecting a battalion node ( X ) with a company node
A fundamental concept here is that, while vision system ( Y ) these probabilities specify:
processing may be complex, with numerous feedback
loops, multiple levels of resolution, recursion, etc., in the P(particular image clusters are companies 1 particular
end we should be able t o associate a deductive chain of region i s battalion)

1340 PROCEEDINGSOF THE IEEE, VOL. 76, NO. IO, OCTOBER 1988
 Rparticular image clusters are companies 1 particular
region i s false alarm)
Rparticular image clusters are false alarms I particular
region i s battalion)
P(particu1ar image clusters are false alarms 1 particular
region i s false alarm).
These conditional probabilities form the elements of a
fi"'
matrix, M ( Y ( X ) , where the (i, j ) t h entry specifies P ( Y j I X i ) .
Messages passed between Bayes nodes consist of prob-
abilityvectors providing evidential support (child to parent)
or model-driven support (parent to child). Evidencecoming
into a node will normally be in the form: P(image features
Fig. 20. Messages out of Bayes node X.
Ey2
I
are due to Y Y ) .
Evidence for a Bayes company node, for example, would
therefore consist of the message: [P(image features are due equilibrium is guaranteed to be reached in time propor-
to a company I company), P(image features are due to false tional to the network diameter.
alarms 1 false alarm)]. Pearl specified his propagation algorithm for static
Amodel-driven support messagefrom anodeXto itschild hypothesis spaces. His results have been extended for
Y consists of a vector derived by dividing (component-wise) dynamic hypotheses in order to apply to the ADRIES appli-
the current belief vector by the current evidential message cation. The Bayes net evolves dynamically during process-
from Y to X. ing. New hypotheses get generated for aggregation of lower
Upon receiving a new message from a parent or child, a level hypotheses into higher level hypotheses, while refine-
Bayes node, X, (for single parent trees) computes its overall ment stages generate more possibilities for a hypothesis.
evidential support vector, Dx, by multiplying (component- Conflicting hypotheses are then merged into single Bayes
wise) all evidential vectors, Dx, from all its children Yi. The nodes.An exampleof thegeneration of a Bayes net i s shown
new resulting belief vector i s the result of: Dx * [(transpose in Fig. 21. This figure shows the generation of two conflict-
Mwx) o Cw], where * indicates component-wise multipli-
cation, o indicates matrix multiplication, Cw i s the model-
driven support message from the node's parent W, and Mwx BATTALIONHYPOTHESES
i s the conditional probability matrix from Wto X. New mes-
sages are then generated for each of the node's children
Yi, by dividing (component-wise) the belief vector by the
COMPANY HYPOTHESES
current evidential message from Yi to X. A new message is
generated for the node's parent W, by computing: Mwx o
Dx, where Mwx i s the conditional probability matrix from BAYES NET
W to X and o indicates matrix multiplication (See Figs. 19
and 20). A is battalion LL B is not battaliin
A is not battalion LL B is battalion
A is not battalion 8 B is not banalion

CONDITIO"
PROBAIILITY
MATRIX

C 8s company h C z 8s company h C j s company h C , IS company

C; s compMy h C, is company h c3 s company h C , s not company
,
C Is company h C2 IS company h C3 8s not compsny h C, IS company
,
c ,scompany h C IS company h C3 IS not company h C, is not company
Fig. 21. Bayes net generation example.
b
fyl %y2

ing battalion hypotheses based on four company

Fig. 19. Messages into Bayes node X.
hypotheses. The diagram at the top shows formation region
constraints of possible battalions (the triangle and rect-
The advantage of using Pearl's algorithm for belief prop- angle) fitted to company formation constraints (the small
agation is that it can be implemented as an active network rectangles). The structure of the Bayes net parallels the
of primitive, identical, and autonomous processors. The structure of the Force Models Database. As larger force
primitive processors are simple and repetitive and thus can hypotheses are generated, Bayes nodes corresponding to
be implemented in a distributed environment. Each node force models further up the part-of hierarchy will be gen-
in the Bayes net can be assigned a virtual processor that erated. As the force type and deployment type of the
sends messages in the form of probabilistic belief vectors hypothesesarerefined, Bayes nodescorresponding to force
to parent or children nodes. When a virtual processor models further down the is-a hierarchy will be generated.
receives a new message, it updates its belief vector and Here we give a simple example of Bayesian inference in
propagates the change to its other relatives. Therefore, ADRIES. The approach is detailed in [44]. Fig. 22(a) shows

FRANKLIN et al.: EXPERT SYSTEM TECHNOLOGY FOR THE MILITARY

1341
 tb,
Fig. 22. Original imagery (a) Original SAR image. rb, Fig. 23. Focus of attention and detection processing. (a)
Ground truth lor beliic le\ Focus of attention region cued trom other intelligence
sources. (b) Vehicle detections

the original SAK image. Fig. 22(b) shows the original image
with ground truth of tank locations indicated. Note that
there are three tank companies, deployed in defensive for-
mations. Fig. 23(a) shows a focus of attention sub-region of
the original image obtained from other intelligence source
cues. Fig. 23(b) shows the results of detection algorithms
run on the focus of attention image region in Fig. 2Ya).
Fig. 24(a) shows the results of applying the defensive for-
mation tank company mode matching algorithm to the
detections. This i s the model of Fig. 15(b). Note that match-
ing was restricted to a localized search area, rather than pro-
cessing detections from the entire image. Because partial
matches are allowed, many incorrect and partially correct
matches are made. Because more than one company can-
not claim the same vehicles or occupy the same territory,
each of these tank companies i s in conflict with the others.
Fig. 24(b) shows the resolution of the conflicting matches.
Fig. 25(a) shows the probabilities involved in resolving
conflicts between competing company hypotheses. The
notation "S< three-digit-number >" indicates a grouping
oi three tanks in a line with approximate spacing corre-
sponding to squad groupings in the defensive deployment
formation being searched for by the tank company match-
ing algorithm. The notation "C<two-digit-number >" i s the
grouping into companies formed in the matching process.
Note how multiple companies attempt to claim the same
tanks grouped in squads. The numbers to the side of the Fig. 24. Tank-company matches and resolution. [a) Con-
flicted company matche5. &) Kerolved inatc hes
Csaretheprobabilitiescomputed in supportiorthesecom-
pany hypotheses based on the calculation in [44].
Fig. 25(b) shows the match of the doctrinal battalion At the current state of system development, ADKtES suC-
defensive tormation tothecompany data. Note thatthefor- cessfully infers the presence and locations of military units
mation structure predicts the location of the (not vet found) through the regimental level in high false atarm data, dern-
third company and bat?alion heady uarters. onstrating the ieasibility of the approach. Knowledge is

1.342
 ably limits performance. Machine learning i s an emerging
c25 c22 technology that can potentially solve a significant port ion
oi' this problem by making systems self-modifying, which
618 will improve performance. In expert systems, the improve-
c:3 ment of rules in the system can lead 10an improvement in
C14 system performance.
C8
c15
s354 B. The Machine Learning Approach
(ai Given that the objective of learning i s to improve per-
tormance, several design constraints can be placed on any
system that claims to learn. First, for a system to improve
its future performance in a nonrandom fashion it must, at
some level, evaluate its current and past performance. The
evaluation or critique must consist of comparing observ-
able behavior with some desired behavior. The problem of
evaluation is referred to as apportionment of credit (611, or
credit assignment [62].The apportionment of credit i s to
those etements of the system responsible for good per-
formance. The diagram in Fig. 26 shows the basic teedback

Fig. 25. P r o b a b i l i c t i c reasoning for conflict resolution. ict)

P r o bab i I it i es ~ S S O Ciated to co111piinic's and con $1 i ct ttrf squads.
(bj Partial h t t a l i c t n match.

being atltled in i\C>KIES to accomplish reasoning about s i t -

uation deployiment of toIces in terrain. This will enable Fig. 26. General feedback mechanisin.
A[>KIE:S to predict deployment variation to accommodate
terrain constraints. Capability for software simulation of
input data i s being developed to make statistically signii- mechanism, critic, and apportionment needed to evaluate
icant testing o f .ADRIIIS pcrssilde, performance. In the context of an expert system, the appor-
The research for AIXIES was originally supported by the tionment of credit problem boils down to rewarding good
Deferise Ad v a nc: ed Itesea rc h f'r oj txt s Agency (DAKPA), by rules and (possibly) punishing bad rules. Secondly, theeval-
the U.5. Army intelligence Center and School (USAICS) uation should lead to or support some lasting modification
under U.S. Cover n men t Cont raci DAEAI 8-83-C-0026, and of the system so that the system avoids the same mistakes
by Air Force Wright Aeronautical Laboratories IAFWAL) nexttime. Again, for expert systems, this requires that a sys-
under U.S. Government Contract F~:36.1.~-83-(1:-1070.It is tem be able to alter its rules on the basis of the evaluation.
currently supported by DAKPA and the U.S. Army Engineer The rest of the section on learning describes research at
To pograph ic Labor ai o r ies (C1 SAE.TL) u nde r U S. Govern- I
the MITRE Corporation, a n attempt to apply production sys-
ment COnt riict DACX76-86-C -0010. AI3R 11. S i s the joint wo rk tem technology to p r o b l e m in military intelligence. The
o i manv rc?warchersat iidvant-etf Ilecision Systems (ADS) difficulties associated with knowledge engineering of a
i n hlou nt ai n View CA, S c ie n c: e Applications In t er nat ior ia I
I
large-scaleexpert system (e.g., extended knowledge acqui-
COr porat ion iS A i C1) in 'I u(:son,A%, 3 t i cl t 1.1e iLZ RJ Cor por at ion sition, preference-biased expert knowledge, static knowl-
in Claktcrn, VA..Thc project supervisors and principal inves- edge in the face of adaptive adversaries) led to the consid-
tigators are Rob Drazovich and Toci l.evitt at ADS, Dick eration of incorporating learning capabilities.
Kruger and I.arrv '&inter ai SAIC:, and K o h Ready and Chris Learning systems are usuatlycharacterized by three basic
MCKW ijt M R \ . etements: domain of application, knowledge representa-
tion,and learningstrategy. Sections D, L a n d Fdiscussthese
elements in the context of an existing implemented system
A. The ."ccd f o r Lcxnjng called M2. 'The discussion concludes with issues and f u r -
ther research topics. This brief overview of the basic
The previous sections have dernonstrated the expert sys- approaches to machine learning is not intended to be a
tems a p p r o i ~ lto i computer-based problem solving. This tutorial o n learning; rather, it serves to place M2's approach
approach i s applicable to a wldc~variety of problems and in the context of previous research.
provides a C O nven i e s nt mE:(: han i s rn for expressing k nowl-
edge. lirrfortunately, there Are difticulties associated with
C. Machine Learning Paradigms
t h e approach t h a t limit i t s intrinsic usefulness. The inability
of expert systems to adapt to new situations outside of a A number of approaches to incorporating learning into
part icu Ia r s p 1.1E: r o ik now1ed g e i f rey uent Iy te rined Brittle- computer systems have been developed since the 1050s.
ness or iallir-zg off rhe k/?nwledgt. (:/iff 1\59], [ti011 unaccept- Michalski [63] has identified three basic paradigms in the

13-13
machine learning field that include the learning of concepts If I believe that the goal of UNIT 1 is to move UNIT 2 to
(with an extensive domain model and without one), and RIVER 1 at TIME t, then I expect UNIT 1 to move his lead
learning by self-organization. elements to RIVER 1at TIME t - 2. Further, if I put a sensor
Acquiring Concepts: Symbolic concept acquisition, SCA, there, I expect a report at TIME t - 2 or TIME t - 1.
attempts to formulate general concepts from specific exam-
In general, no information is received which directly cor-
ples. These systems have a teacher that provides explicit
roborates the hypotheses made by the analyst. The only
positive and negative instances of a target concept. In this
feedback available i s in the substantiation of (hypothesis
case the feedback mechanism i s simplified by the descrip-
generated)expectations bythecontinuing stream of reports
tion of the example as a positive or negative instance of the
emanating from the sensors. The analyst generates the
concept. An example system which makes use of this
hypotheses by applying an underlying model of the ene-
approach is Quinlan’s ID3 program [64]. One application
my’s behavior to the data. There are four basic sources of
of ID3 produces classifications of chess end games using
errors in the analyst’s description that require more than
King-Rook versus King-Knight positions. The approach
the traditional expert systems approach to solve: 1) noisy
shows significant speed improvement over more tradi-
and incomplete data, 2) incorrect models of enemy behav-
tional search procedures like minimaxand discovered win-
ior, 3) deceptive enemy behavior, and 4) adaptive enemy
ning strategies overlooked by expert human players.
behavior. Arguably, problems 2 and 3 could be solved with
Using Domain Models: A second paradigm, referred to
enough knowledge engineering, if deception i s considered
as knowledge-intensive domain-dependent learning, KDL,
to be another set of rules to be captured.
uses large amounts of domain-specific knowledge to build
These domain considerations lead to some projections
a model of the domain from which modifications can pro-
about the capabilities of a learning system performing the
ceed. The primary difference between this and the SCA
analyst’s task. First, it must generate its description without
approach is intheamountof knowledgeused bythesystem
access to ground truth. This requirement i s representative
to formulate useful concepts and their frequent use of sym-
of the self-organizing class of learning systems; however,
bolic logic and theorem-proving approaches to develop
the strategy employed is one of inductive learning in gen-
consistent concepts. A well-known system which discov-
eral. See Section F for more on learning strategies. The ana-
ered interesting mathematical concepts using this approach
lyst never knows absolutely whether his analysis i s correct
was Lenat’s A M system [65]. In AM, the search through the
even when his hypotheses are substantiated. Second, the
space of possible concepts is guided by a utility function
system will operate in noisy and possibly discontinuous
with multiple objectives defining the interestingness of the
search spaces in which behaviors are incompletely or
concept being explored. In general, the KDL approach has
improperly executed. In general, this requirement most
an applications orientation.
closely reflects the capabilities demonstrated by research
Building from the Ground Up: The last paradigm, termed
in the self-organizing approach. Finally, the system should
self-organizing systems, often exploits sampling-based
take advantage of the existing organization and models of
learning algorithms rather than symbolic logic approaches
the domain, rather than learning from the ground up (an
to improve performance[66], [671. These systems use a series
argument for the KDL approach). The existence of a model
of evaluation functions and a direct feedback mechanism
is a good reason for not starting the search (for good rules)
to the rules, nodes, or networks which determine the
from scratch. Any reliable search strategy will do better
behavior of the system. An example of this approach is the
given a starting point closer to the objective, and it is
work done in classifier systems and genetic algorithms by
assumed that existing models are fairly accurate to begin
Goldberg [68]. This system was given a random set of rules
with. The existence of a domain model points up a need for
and tasked to perform agas pipelining problem. The system
being able to construct and manipulate high-level data
developed rules to efficiently route and control gas flow in
structures like production rules. The structures with which
the network. The genetic approach is discussed in more
M2 stores and manipulates the model are the topic of the
detail in Section F. As a matter of perspective, it should be
next section.
noted that these paradigms are not mutually exclusive and
that opportunities exist for cooperative learning efforts.
E. Knowledge Representation
The knowledge in M2 is represented in two forms: the
D. The Military Intelligence Problem
“declarative” knowledge and vocabulary of the system, in
The problem associated with intelligence analysis is char- our case an object description and datatype language,
acteristic of a broad class of problems associated with sub- expressed as a taxonomy of the terms of interest; and the
stantiating hypotheses given a model and some data. Con- “assertional” dynamic knowledge of the system expressed
sider a set of sensors which provide the intelligence analyst by pattern-matching production rules.
with a data stream of reports of enemy activity at varying Fact Bases: In M2, objects in the domain are stored in a
time intervals. The task of the analyst is to generate a frame representation language. Theclasses of other related
description oftheenemyunitsthat aregeneratingtheactiv- objects help organize the description by characterizing the
ities detected by the sensors. This description consists of object in terms of its class membership as well as its prop-
identities (e.g., UNIT I), locations (e.g., RIVER1 at TIME t), erties and their values [69]. An example of an object rep-
and goals (e.g., to have UNIT 1 MOVE UNIT 2 to RIVER 1 at resented in this way appears in Fig. 27. The diagram shows
TIME t). The description leads to expectations of further the attributes of the military unit, UNIT 1, and the values
reports and suggests experiments (sensor tasks) which the of those attributes. The values of the attributes are restricted
analyst should execute. An example of expectations would by the properties of the attributes as with the ECHELON
be: attribute in the figure. At this point, the structure of the

1344 PROCEEDINGS OF THE IEEE, VOL. 76, NO. IO, OCTOBER 1988
 Echelon
1 Unit 1 the inference by generating plausible hypotheses and

I
selecting among them.)
domain units echelon (corps) M2 Architecture: Previous sections have attempted t o
range instances commander smith
instances personnel 10000 describe the constraints under which a learning system in
dlV equipment armor the intelligence domain must act. The overall picture of the
bde engr
bn arty flow of information between the system modules i s shown
CO location valley1 in Fig. 29. The system receives input from the user in the

Fig. 27. Object representation.

I Inference I

objects of the domain are static and not subjected t o the

learning strategy. This was an initial design decision t o focus
o n the learning of behaviors rather than concepts, because ou al
the objects of the domain are assumed to be well known.
Rule Bases: The behaviors associated with objects, and
the relationships between object behaviors and attributes,
and object goals are stored and manipulated in production
rules. An example of a production rule for recognizing a
river crossing goal appears i n Fig. 28. The translation of the
I
User Input
1: Facts about analyst
2: Analyst model of planner
IF (unit-is ?name ?equipment Vocation) 3: Facts about planner
(river-is 7riv-name ?location 1 )
(within-distance ?location ?locationl)
(includes ?equipment engr) Fig. 29. Domain driven architecture.
THEN
(goal-is ?name cross-river 7locationl)
Fig. 28. Rule representation. form of a goal for the planner to achieve. The planner, with-
out access t o the analyst’s knowledge bases, generates a
rule is: if there i s a unit with engineering assets near a river plan t o achieve the goal and sends the plan t o the simu-
then its goal is t o cross the river. It should be noted that the lation.The simulationcarriesouttheactionsofthe plan and
generates activity assertions that are available t o the ana-
clauses of the rule in the diagram correspond t o objects
lyst’s sensor rules. Sensor rules generate reports based o n
defined elsewhere. The clauses, frequently called relations,
their own coverage (in time and space) of the activities. The
have the same internal structure as their object counter-
reports are used by rules t o generate a description of the
parts. Thus, the datatypes associated with the fields of each
planner (as described above), and predictions of future
relation are restricted i n the same waythat attributes of the
reports are made. The degree of correlation between these
objects are restricted. The details of these datatype restric-
reports and incoming reports prompts the learner t o mod-
tions become important when the learning mechanism
ify the analyst’s knowledge base of the planner. The pri-
begins creating rules of i t s own. We want the learner t o
mary points of interest in the architecture are: 1)separation
explore the space of legal rules rather than the greater set
of planner and analyst knowledge bases, the ground truth
of all possible combinations of syntactic elements. The pro-
restriction, 2) user input t o planner i n the form of some goal
duction rules that can be generated by combining relations
to be achieved, and 3) learning driven by the output of the
define the space of possible rules t o be searched by the
analyst t o modify the analyst model of the planner.
learning strategy. Obviously, a simple enumeration of all
1earningModule:The learning module is divided into two
the legal rules is impractical since the number of rules var-
distinct components oriented around the evaluation and
ies exponentially with the number of relations. The next
modification of knowledge. Evaluation results drive the
section describes a procedure for searching the space of
modification of knowledge. We begin by discussing M2’s
legal rules efficiently, focusing o n areas that need improve-
apportionment of credit mechanism, avariant of the bucket
ment.
brigade algorithm [70], which was reimplemented in the
context of our rule representation and i n an OPS-like [71]
F. Strategies for Learning
inference mechanism.
A strategy i s a plan or series of plans whose actions Bucket Brigade a n d Apportionment of Credit: A com-
accomplish a goal. A learning strategy then, is that set of plete description of the bucket brigade implementation i n
actions or the methodology that will be invoked t o improve this system and results are presented in [72]; however, i t w i l l
performance. The general types of strategies available are: be useful t o review the ideas presented there. As we indi-
1) learning by rote, 2) learning by instruction, 3) learning by cated in Section E, M2’s rule clauses contain either variable
deduction, 4) learning by analogy, and 5 ) learning by induc- or literal values i n their fields. Let f, represent the total num-
tion [63]. Although a f u l l description of these learning strat- ber of fields contained in the rule i’s clauses, and v, rep-
egies i s beyond the scopeof this paper, it is useful t o notice resent the number of fields which have variable values. A
that the learning strategy defines the search mechanism heuristic measure of specificity is computed, s,, for each
employed in the space of rules. The strategy of learning by rule as
induction i s used because it closelymodels whatthe human
analyst must do. (A strategy is inductive when it performs s, = 1.0 - v;/f,.

FRANKLIN et al.: EXPERT SYSTEM TECHNOLOGY FOR THE MILITARY 1345

Specificity, as computed, acts as a measure of match for a
rule to a set of short-term memory items that are matched
by the rule’s conditions. Strength will be used as a record
of a rule’s performance. Strength and specificity tradition-
ally play a part in the bucket brigade serving as objectives
in a utility function known as the bid. Let S,(t)represent the
current strength of rule i at time t. A constant c represents Hypotheses\
a moderator much less than 1. Let the bid, B,(t), of a rule
at time t equal the product of c, s,, and S,(t): Bid Payment\ Rule
U
B,(t) = c * s, * S,(t).
Observably correct hypotheses
The bid is used to produce a linear ordering of rules in the
conflict set during the conflict resolution cycle of opera-
ENVpayofl
tion. Rules which support rule i a t time twill be represented
by supp,(t). Finally, ENV will be used to denote the envi- Fig. 30. Bucket brigade process.
ronment. It has an attribute, ENVpaYoff, which represents the
amount of reward the environment provides to rules (typ-
right sets of relations and restrictions in the first place. This
ically an order of magnitude greater than the average inter-
points out the need for structural modification of the rules.
rule payoffs described below)whose predictions of observ-
Modification Operators: The performance of rules in the
ables exactly match observables from ENV.
system are reflected by the rule strengths and the fre-
The somewhat lengthy definitions above help t o present
quencywith which the rules are used. (A rule that has never
succinctly the operation of the bucket brigade algorithm
matched and never fired can hardly be evaluated beyond
adapted to the rule representation. Rules that match items
i t s general inapplicability.) The rule base in M 2 is divided
in short term memory are entered into the conflict set. Each
into rule sets whose member rules make the same class of
rule in the set submits i t s (closed) bid, B,(t). All rules in the
inference. For example, rules that reason about equipment
set are then ordered based upon their respective bids. The
behaviors are separated from those that reason about bat-
highest bidder i s selected for firing, i s fired, and the bid
talion behaviors. If one considers the operation of the
made bythat rule is distributed equallyto supp,(t) and added
bucket brigade, it can be seen that a differential in strength
to their strengths Sj(t), {for j an element of supp,(t)}. Con-
from the beginning of the reasoning chain t o the end devel-
versely, this bid is subtracted from S,(t)of thewinning rule.
ops. However, within a single class of inference any dif-
This primitive rule economy leads t o reinforcement of rules
ference reflects more of the relative merits of the rules and
on the basis of their utility t o the system.
less their position in a chain of inference.
At this point, the flow of strength in a chain of rules i s
This research takes advantage of the strong attributes of
from last t o first. The last rule in a chain has no mechanism
inductive approaches to learning developed by Holland and
for regaining its bid, leading to a gradual weakening of all
the genetic algorithms community. This initial approach
the rules in the chain. This problem can be alleviated by
does not preclude the useof other learning techniques nor
ensuring that the rule base i s capable of generating asser-
does it provide commentaryontheir efficacy in the domain;
tions which can be matched against observables in the sys-
rather it is a starting point. The primary effort has been in
tem. The only observables of the analyst module are the
the adaptation of genetic modification operators t o the M 2
continuing reports from the simulation. Our implemen-
representation language [73]. Although genetic algorithms
tation employs a set of meta-level rules with the general
have been implemented in n-ary languages, they have not
pattern:
been implemented in systems employing rigid datatyping
IF and hierarchically organized object and behavior spaces.
(there i s a report of some kind) When To Modify?:The indications for modifications come
and in two significantly different forms. The first, cataclysmic
(there is a prediction of that report) failure on the part of the system, indicates a lack of the rule
THEN sets (or possibly the entire rule base) t o lead to useful and
(payoff the prediction rule with ENVpaYoR), correct predictions about the environment. Ordinarily, ran-
domly generated rules would not exhibit such failures
as the agents for effecting this reinforcement of the pre- unless a drastic shift in the application domain of the sys-
dictions. These rules take the place of the critic in Fig. 26. tem had just been undertaken. A second, less disheart-
The behavior of the bucket brigade is greatly influenced ening indicator i s thegradual differentials which arise in the
by the value of c. The constant acts as a moderator and strengths of rules at the same reasoning level. An exarni-
restricts rules to risking onlya small fraction of their wealth nation of theconditions underwhich this differential arises
on any cycle. This acts as a mechanism for reducing the would suggest the appropriate when and how of modifi-
chancesof overfittingthedata.The basic procedure behind cation. The M 2 system assigns a threshold to the standard
the bucket brigade is shown in Fig. 30. The form of strength deviation of rule strengths in a rule set. Passing this thresh-
revision in the bucket brigade is a form of learning because old serves to direct the attention of the modification oper-
it has lasting effects on the performance of the system, i.e., ators to the offending rule set. One of two things could be
rules that have done well have an increasing advantage for true in this situation: 1) some rules are winning a lot and
firing over rules that performed poorly. However, strength collecting all the payoff, or 2) some rules are losing a lot and
revision may not be adequate if the rules do not contain the losing all of their strength. Whether the rules were winning

1346 PROCEEDINGS OF THE IEEE, VOL. 76, NO. IO, OCTOBER 1988
or losingthey must have been firing,which meansthey must straintsoftherule,WITHIN-DISTANCEand INCLUDES, may
have been matching situations i n the system. The strength be applicable in a broad class of rules about crossing
and frequencyof firing depends in part o n thetypesof rules objects, and the fitness proportionate reproduction and
extant in the rule set. The t w o different varieties of rules i n recombination methods would promote the use of these
rule sets are: 1) specialist rules matching specific sets of constraints.
assertions, and 2) generalists, that can match a variety of The experimentation into the behavior of the modifica-
assertions. The learning task is to discover the proper mix tion operators in M2 is continuing. Current work i s cen-
of generalists and specialists (notice the bias i n the bidding tering o n strengthening the theoretical foundations of the
functions toward specific rules) capable of producing good recombinant approach i n the higher-level representations
ove ral I perform ance. M2 employs. A series of experiments to determine relative
How Should I Modify?: The M2 system contains three measures of performance between the t w o approaches to
basic kinds of modification operators: specializers, gen- modification i s planned as of this writing.
eralizers, and recombiners. In terms of the pattern-matched
rules discussed above, specialization amounts to either G. Issues and Future Topics
changing variable values to literals, or adding constraints
Avarietyof technical issueswith respect totheM2system
o n variable values. Generalization, o n the other hand, is
exist. In terms of evaluation, the rate of learning i s limited
exactly the opposite. Generalization operators change lit-
by the number of rules in a chain of inference before an
erals t o variables and remove constraints [73]. Using the
ENVpayo~ i s received. As implemented, the system requires,
river-crossing rule in Fig. 28, a generalization would be t o
as a lower boundary, a number of iterations equal t o the
remove the constraints (WITHIN-DISTANCE ?LOCATION
lengthofthechain beforethestage-setting rulesatthe head
?LOCATIONl)and (INCLUDES ?EQUIPMENTENGR).A spe-
of the chain receive a payoff. Holland has suggested [70]
cialization would betochange?NAMEto UNITI. Iftheoper-
(and Riolo implemented) [74] a remedy for this problem; the
ators strictly modify the rules, then the orientation of these
implementation in M2's representation remains as a future
operators, as expressed, i s completely different than that
topic. A second issue related t o evaluation is the use of the
of the bucket brigade and the stated purpose of our mod-
utility measure as a strength revision mechanism. Notice
ifications. Note that a strict replacement scheme tries to
that the system as described selects for useful rules rather
generate the best rule for performing in the environment,
than strictly correct rules.
rather than generate an increasingly stratified (in terms of
Some of the most creative work in the field of machine
specificity and generality) population of rules t o produce
learning is in the development of useful modification oper-
better performance. The incorporation of operator modi-
ators. A current debate in the field is the relative merits of
fied copies of these rules can produce the desired effect at
the logic-based approach compared t o the sampling-based
the expense of an initiallyexpanding rule base. The remain-
approaches we have described. Future research will focus
ing requirement i s that these operators have some con-
o n this issue particularly in the context of the required
ditions for use. The utilization of meta-rules in learninggen-
knowledge representations for each approach and the con-
eralizations and specializations is one of M2's approaches
straints those representations impose.
t o modification. An example of a (rabid) generalizer meta-
This section deals with expert systems as primarily nota-
rule in M 2 is shown below:
tional, problem solving conventions rather than as models
IF for the human cognitive process. The area of learning
(rule-is ?name ?rule-set ?strength ?frequency ?spec- research concerned with modeling human learning [75] is
ificity) not treated here. Cognitive psychologists have made sig-
(rule-set-is ?rule-set ?ave-strength ?ave-specificity) nificant advances toward building systems that model
(exceeds-threshold ?strength ?ave-strength) human learning phenomena [76]. Discoveries i n these
(specific-rule ?specificity ?ave-specificity) endeavors have led t o formal theories of general learning
THEN mechanisms.
(associate ?new-rule (make-copy ?name)) Machine learning i s a field i n its infancy. Many very dif-
(remove-predicate-constraints ?new-rule) ficult problems remain unsolved including issues related
(variablize-literals ?new-rule). t o the rate of learning, overfitting of data, and general prob-
lems related to efficient search and inference. A final issue
This meta-rule makes a copy of the rules that satisfy its con-
related t o knowledge representation is the use of the most
ditions, removes the constraints in the new rule, and
efficient versus the most transparent representation. The
changes its literal values to variables.
performance of learning techniques i n large-scale prob-
Preservation of rules and parts of rules i s a natural ele-
lems will determine the efficacyof approaches being devel-
ment of the recombination operators i n M2. The recom-
oped.
bination operators make use of notions from natural
genetics and were first described in thecontext of machine
V. THE NEXTSTEP: DISTRIBUTED
PROBLEMSOLVING
problem solving by Holland [67]. The idea behind the
recombination of rules is that through fitness proportion- The potential benefit of systems such as BATFLE, ADRIES,
ate replication of rules and the recombination of the repli- and M 2 is extensive. It i s unclear, however, how the poten-
cants with the original population, the average perfor- tial utility of these and other expert systems will be trans-
mance of the rule population will improve. The parts of the ferred t o real applications in operational environments. Test
good rule are extant in the population at a greater fre- methodology, configuration management, reliability,
quency and thus appear o n average i n more rules. For maintainability, performance, and system security are just
example, consider the river-crossing rule above. The con- some of the practical issues that must be resolved realis-

FRANKLIN et al.: EXPERT SYSTEM TECHNOLOGY FOR THE MILITARY 1347

 linked via communication paths of various types). The sys-
tically for military expert systems. Failure to fully address tem as a whole may be both data-driven and goal-driven,
these issues will probably cause the early (if not preemp- by permitting each expert node to select different cate-
tive) retirement of many future systems. gories of work dynamically.
In addition, several important military problems are not Much of thecurrent research into DPSS involves systems
suitable for conventional expert system approaches, even descended from the HEARSAY experiments [80]. The HEAR-
though this technology has been applied to many diverse SAY family of expert systems started with HEARSAY-I, an
domains, including medicine, geology, mathematics, man- earlyexperiment in speech understanding [81]. After HEAR-
ufacturing, finance, and education. For most of the suc- SAY-I, evolution continued with HEARSAY-II, which
cessful applications, however, at least one common char- improved the flexibility of both knowledge representations
acteristic can be observed: solving a typical problem and problem-solving control strategies [82], and HEARSAY-
requires knowledge only about a single, bounded domain. Ill, which provided problem domain independence [83].
For example, a system in a hospital that makes diagnoses The HEARSAY paradigm involved the processing of input
about patient illnesses can reach acceptable conclusions data through multiple levels of detail using a blackboard
without having any knowledge of hospital staffing prob- data structure (blackboards are discussed in more depth in
lems or accounting procedures. the next section).
This same characteristic i s also a good reason why expert The use of multiple levels allowed the modularization of
systems have not been fully exploited in certain military expertise for each level, thereby permitting each module
applications; many critical military problems require a towork in a limited domain and apply contextual cuesavail-
broad perspective that spans a variety of domains. For ableatthat level. For instance, referringto Fig. 31,thelowest
example, the problem of determining a country’s inten-
tions based on intelligence about its local troop movements
could require the effective application of knowledge about 1 LEVELS I INFORMATION I
many different complex domains: the local tactical situa- Complete Sentence
Data Base
tion, the history of that country’s deployment patterns, the Interface

force posture of its allies and neighbors, the political and Inferred Phrases
Phrase nllllllllllllllllllllilllllllllllll
economic situation (both within that country as well as
between it and others), the public pronouncements of that Inferred Word Sequences

-- --
country and others, the current weather, the capabilities of ,:Zke
mzza mzzz
collection systems, and the proper interpretation of spe-
Inferred Words
cific types of collected intelligence, including [MINT and

- -
Word
SIGINT. Since no one human expert could master this
Inferred Syllables
breadth of knowledge, it is reasonableto conclude that this Syllable 000 00
problem will not be solved soon by any single, monolithic
expert system, either. In fact, some automated systems
(such as TRICERO [77l and ALLIES [78]) have achieved mod-
Segment .-
Segmented Signal Data
Y Y Y Y U-

1
est successwith thistypeof problem solving by distributing
the work among multiple cooperating expert systems. 1 Parameter 1 Raw Signal Data
I
What is required is an open network architecture thatwill
Fig. 31. Construction of higher-level patternsfrom raw data
permit multiple expert systems to communicate and coop- in HEARSAY processing.
erate effectively and coherently [79] on the solutions to
large, complex problems. Numerous advantages may be
realized by distributing a large expert system problem level of HEARSAY divided the raw input data intosegments;
across a set of cooperating experts, including the extension the next higher level collected segments intosyllables, then
of a system across multiple physical systems, greater sim- into words, word sequences, and so on. The actual perfor-
plicityof each component expert, and the use of parallelism mance involved the creation of “islands” of information on
to speed up the resultant system. This technique can be the blackboard where HEARSAY could make inferences of
applied quite naturally to some of the more intractable mil- high confidence. These islands were extended to include
itaryanalysis problems, since it mirrors thegroupsof human neighboring information and cues until enough informa-
experts that work as units within large military organiza- tion was correlated to create an inference at the next higher
tions (e.g., an intelligence analysis team or a battle planning level.
team). This process of dividing the inference structure into mul-
tiple levels of detail can be applied directly to large, com-
plex military problems. Consider, for example, the simpli-
A. Distributed Problem Solving Systems
fied I&W process depicted in Fig. 32. Raw sensor data is
The concept of a distributed problem solving system collected into messages or processed directly; messages
(DPSS) can be defined as a loosely coupled network of are collected into indicators; and indicators are processed
expert system nodes that communicate both information into warnings, which are sent out of the system or fed as
and goals to each other in the process of arriving at com- input to other applications. The output of such a system
mon solutions. A global blackboard(discussed later) serves consists of the patterns recognized in the dataand message
as the logical communications medium. The experts to be stream by the set of experts who make up the processing
coordinated may be both logically and physically distrib- power of the system.
uted (i.e., some experts may be co-resident on one machine Thus the HEARSAY blackboard paradigm i s appropriate
while others may be distributed over several machines fortwo reasons. First, it solves a problem that is structurally

1348 PROCEEDINGS OF THE IEEE, VOL. 76, NO. IO, OCTOBER 1988
 -
LEVELS INFORMATION
nodes will generally have numerous goals to choose from,
and will use some criteria for selecting the most valuable
Predictions goal at a particular time.
The final type of information associated with a DPSS
1 1 Warnings
blackboard is derived from i t s distributed nature. This
information i s required t o control the flow of information
Assessments from one expert’s blackboard t o those of other experts. In
a single-node expert system using a blackboard architec-
ture, one blackboard will be the central repository of all
information about the problem. With a DPSS, however, each
separate expert node requires access t o blackboard infor-
Matching
, , I , I mation. This implies that expert nodes should use local
blackboards that are actually individual partitions of an
abstract global blackboard. One or more strategies can be
employed by each expert t o select the information that
should betransmittedfrom or accepted intothe local black-
board, the other experts that should participate in an infor-
Fig. 32. Construction of higher-level patterns from raw data mation exchange, and the conditions under which such
in I&W analysis. exchanges are desirable. The alternative i s a single black-
board maintained as a service for all of the experts. How-
ever, this technique has severedrawbackswhen theexperts
similar in nature to intelligence data analysis problems, as
are physically distributed. Since the blackboard i s a vital
well as to other complex military problems requiring the
resource to the inference process, thecommunications load
coherent integration of diverse expertise, such as battle
involved in supporting a centralized blackboard would be
management. Second, as discussed in the next section, the
immense. I n addition, a single global blackboard would
blackboard is an especially appropriate control structure
provide a single point of failure in a physically distributed
for DPSS.
DPSS.
Thus, each expert node of a networked DPSS should have
B. Blackboards in DPSS
a local blackboard composed of information directly avail-
A blackboard is a global data structure used t o com- able t o that expert. Some of this information may be irrel-
municate facts, inferences, and goals between multiple evanttothatexpert,and someof it may be required byother
expert nodes in a DPSS. Its power results from the way it experts. During the course of problem solving this “mis-
provides a common framework and interface for sharing placed” information will be communicated from one black-
useful information both within and among expert nodes. board to another according t o blackboard communications
Its flexibility supports both data-driven and goal-driven criteria. For instance, when a particular confidence thresh-
operation. old was reached in an inference generated by one expert,
Three major types of information are represented in the it might be “leaked” to a sibling or higher-level expert for
full blackboard paradigm. The first type is frequently further processing. This process of leakage i s detailed in
referred t o as “data,” but it can be further divided into facts Fig. 33. In the case of information needed by the receiving
and inferences. A fact i s some datum that has been inserted expert, it will be used as an inference t o drive further pro-
into the system from an external data stream o r a human cessing. I n the case of information received by an expert
user. It may have a confidence factor assigned by (or nodeviathis processand substantiated, a return flowwould
because of) i t s source, but it i s not generated by the system be used to bolster the confidence of that inference in the
itself. By contrast, an inference is produced by the appli- original node. In the case of conflicting information, a
cation of some sort of knowledge to the fact sand inferences reverse return flowwould lowertheconfidenceand/or spur
already available t o the system. The confidence factor asso- error recovery actively i n one or both experts. I n the case
ciated with a given inference i s a function of the system’s of goals propagated between blackboards, the receiving
confidence in the knowledge that produced the inference expert node would have t o choose among externally-sup-
and i n the other facts and inferences used t o support that plied goals and its own goalswhen evaluating which t o exe-
inference. cute [79]. This localized control ability is discussed in more
The second type of information i s “control” information. detail later in this section.
Again, this i s further broken down into knowledge sources The blackboard paradigm thus provides a method of
and goals. Knowledge sources are the procedures that loosely coupling various expert nodes supporting conflict
define how inferences are t o be drawn and actions are to resolution and individual activity. The paradigm is not cen-
be performed by the system during its normal mode of tralized,sothatthere isagracefuldegradation ifoneor more
operation. I n some implementations, knowledge sources nodes are removed from the system (assuming that the loss
are able t o pre-identify both the data required for them t o of one node’s expertise or information does not prohibit
be effective and the types of inferences that they can pro- effective problem solving altogether). The system may be
duce. Goals represent more general activities that may be data-driven as information islands are propagated upward
undertaken by an expert, such as the generation of an infer- to higher-level or sibling nodes, or it may be goal-driven by
ence that matches a certain pattern (if possible). I n some placing goals o n the local blackboards attached t o various
of the literature, a separate data structure known as an nodes. Communication load may betuned by changing the
agenda i s established for the maintenance of goals. Expert rules controlling inter-blackboard information movement.

FRANKLIN et al.: EXPERT SYSTEM TECHNOLOGY FOR THE MILITARY 1349

 Node 1
edge sources it has available, the control decisions it makes,
Local
Blackboards and the communication policies that regulate the transfer
of inferences between experts.
Local Knowledge Source Distribution: I n some DPSS archi-
Processes
tectures, all of the knowledge sources are directly available
(a)
to all of the expert nodes [86], while in others, each expert
has access t o only some knowledge sources [85]. The deter-
Node 1 Node 2 Node 3
mination of which knowledge sources will be available t o
Local
Blackboards each expert i s usually made by the system architect, based
on an evaluation of thedifficultyof hosting different knowl-
Local
Processes edge sources o n different nodes. Knowledge sources that
are initially designed t o be part of a given DPSS are likely
to be easier to share among experts than knowledge sources
Node 1 Node 2 Node 3 that already exist when the DPSS is designed.
Local
Task Distribution: When considered as a monolithic
Blackboards
operation, each knowledge source iscapableof performing
Local one or more units of work or tasks. The potential tasks that
Processes knowledge sources can perform are usually represented as
(C) goals, which define tasks in terms of the possible result pro-
duced by knowledge sources. During DPSS operation, the
Node 1 Node 2 Node 3 set of tasks under consideration at an expert node i s a func-
Local
Blackboards
tion of the inferences produced by i t s own knowledge
sources, the communication policies that regulate the
Local
Processes
transfer of goals between experts, and the degree to which
agiven goal can beeasilyfactored into local sub-goals. Since
(4
any one node can actually work o n only a limited number
Fig. 33. Data “leakage” between local blackboards.(a) Data
of tasks (usually one) at one time, the method by which it
placed on first blackboard not certain enough to reach rep-
lication threshold. (b) Data reaches certainty threshold for evaluates and activates potential tasks has a critical influ-
replication (relativecertainty is indicated by line thickness). ence o n the usefulness of i t s problem solving activities (as
(c)Feedback enhances certaintyof data and reaches thresh- discussed in the next section).
old for replication to next node. (d)Conflict propagates neg- Communication Policy: As already noted, each expert
ative certainty back.
node in a DPSS must implement a communication policy,
which need not be the same for all nodes. From the per-
spectiveofa sending node,this policy determineswhen the
C. DPSS Attributes
node will transmit information (either inferences or goals)
A number of different attributes may be used t o compare to other nodes, what types of information will be trans-
and evaluate alternative DPSS architectures. Each of these mitted, and which nodes are appropriate recipients of the
attributes can have a profound impact o n the performance information. From the perspective of a receiving node, this
or even the viability of a given DPSS architecture with policy determines when the node will accept information,
respect t o a particular problem domain: what types of information will be accepted, which nodes
Physical Distribution: The individual expert nodes of are appropriate originators of the information, and how the
a DPSS can be resident i n the same processor (i.e., a logical accepted information should be evaluated locally. The
distribution) [84], assigned t o individual processors in a choice of a particular set of communication policies may
tightly-coupled parallel system [85], or distributed across be influenced by any predetermined organizational struc-
multiple processorscommunicating through a LAN.Acom- ture for a given DPSS. However, the possible subtle inter-
bination of physical and logical distribution i s possible actionsof different aspects of communication policies make
within a single DPSS. The degree and type of physical dis- it difficult t o establish an opfimal set of policies without
tribution for agiven DPSS i s largelya function of the overall extensive experimentation in real or simulated environ-
goals of the system architects. For example, logical distri- ments [87l.
bution provides a means for flexible experimentation with OrganizationalStructure: It i s possible to bias the eval-
different network topologies while requiring a minimum uation of tasks by each expert node so that it prefers certain
amount of processing and communication hardware. activities over others, regardless of the current state of
Data Distribution: During normal operation, each problem solving.This typeof bias, i n conjunction with com-
expert node in a DPSS will receive a different set of facts munication policies, implements an organizational struc-
about the world and inferences from other experts. (The ture across the DPSS. I n a hierarchical structure, for exam-
utilityof any architecturewhere all experts receivethe same ple, each node accepts tasks from at most one other node,
information is unclear, due to the implicitly redundant thereby permitting higher-level nodes t o control and coor-
activity required at each node.) The set of facts available t o dinate the activities of lower-level nodes. In a flat (com-
each expert node can be influenced by many factors, mittee) structure,each nodecan accepttasksfromanyother
including the spatial distribution of data, the network node, thereby maximizing the availability of a given node’s
topology, and the distribution of knowledge sources services t o the whole network. Of course, many other orga-
throughout the DPSS. The inferences available t o each nizational structures are possible. The combination of an
expert nodearea function of the facts it receives, the knowl- organizational structure and a set of communication pol-

1350 PROCEEDINGS OF THE IEEE, VOL. 76, NO. IO, OCTOBER 1988
icies establishes a distributed problem solving strategy Organizational Significance: Tasks that are important

-
within a DPSS.
Organizational Development: If the DPSS architect pre-
t o an expert because of its organizational bias should
be performed before tasks that are less relevant to its
establishes an organizational structure for the system, then known responsibilities.
no organizational development i s required. Alternatively, Efficiency: If it is possible to estimate effectively the
the DPSS can develop an appropriate organizational struc- cost of performing a task before it i s executed, then
ture as part of i t s initial problem solving activity. This can less expensive tasks may be preferred over more
beachieved by requiring nodesto negotiate(as in the”con- expensiveones. This criterion may beespecially useful
tract net” approach of Smith and Davis [88],[89], an appro- when a given goal can be achieved by more than one
priate task distribution p r i o r t o the start of problem solving. local knowledge source.
Another option is t o permit the DPSS t o alter i t s organi- Goal Satisfaction: Tasks that are themselves su bgoals
zational structure during the course of problem solving, by or supergoals of several other tasks may be preferred
allowing nodes to detect and respond t o strong mismatches over tasks that are less connected with other activities.
between their knowledge of local task requirements and Task Age: If a task has been awaiting activation for a
their knowledge of current organizational responsibilities long time or i s approaching an age cutoff threshold,
[go]. The choice of an appropriate organizational devel- it may be preferred over tasks that were created more
opment scheme for a given DPSS depends o n whether an recently (or vice versa).
optimal Organizational structure: 1) can be predetermined
Thus a high priority request from another expert might take
for every problem-solving situation, 2 ) can be determined
precedence over a local “business as usual” task, or a short,
once for each problem-solving situation, or 3) must change
simpletask mightoverridealongtaskthat i s basedon infor-
during the course of problem-solving activity.
mation of questionable validity.
The possible interplay among these attributes can yield
These focus of attention criteria can be implemented as
complexsystem behavior that is not easytoanticipate. Con-
variable, weighted parameters of the DPSS. The operation
sequently, it i s important for DPSS developers to incor-
of the system can be modified by altering these parameters
porate extensive internal auditing and measuring capabil-
for a single expert or a group of experts. Changing the bal-
ities in their systems, so that they can sensibly test,
ance between the criteria of validity and goal satisfaction,
understand, and modify problem-solving results and strat-
for example, can alter system behavior to become more
egies. In fact, some DPSS environments [84] have been
data-directed or more goal-directed; modifying the inter-
developed primarilyforthe purposeof measuringtheeffec-
nallexternal control parameter can determine whether
tiveness of different problem-solving strategies.
individual nodes are more locally directed or more exter-
Regardless of the specific attributes of a given DPSS, i t s
nally directed.
expert nodes must incorporate a method for determining
The use of these criteria allows the focus of attention to
when a potential task should actually be performed (i.e., for
vary with time. During normal operation, the focus would
determining the focus of attention). Each expert may be
follow a logical “train of thought” with respect to a given
attempting to satisfy multiple goals, and some of these goals
expert. Anomalous situations, however, could be handled
may be generated locally, while others may be supplied by
by shifting the focus t o different types of tasks. Requests
other experts. The mechanism for resolving these com-
from a system user or another expert could be handled as
peting claims implements a local problem-solving strategy
interruptstothetrain ofthought,and critical warningsfrom
within each node.
input data or other experts could be dealt with in an effec-
tive manner. Thus, this approach permits DPSS nodes to
D. The Focus of Attention Mechanism function passively, actively, or interactively, depending o n
dynamically changing problem solving requirements.
The DPSS techniques that have been developed at the
University of Massachusetts [go], [91] include an elaborate
E . An Example
focus of attention mechanism. The technique functions by
giving each expert node the ability t o decide which of i t s As part of its ongoing research and development
possible goals it should pursue (through the activation of program, PRC has been exploring the potential military
its available knowledge sources). These goals include both application of distributed problem solving techniques by
those generated by the expert itself and those generated building a prototype DPSS. This system, called the Com-
by others. The focus of attention i s determined b y first eval- municating, Cooperating Expert Systems (CCES) Frame-
uating potential knowledge source activations against var- work, provides an evolving workbench for testing different
ious criteria, and then by selecting the most highly rated DPSS concepts. A brief overview of this prototype and an
task for execution. These criteria could include, but not be exampleof i t s operation may helpclarify someof thesecon-
limited to, the following: cepts.
The CCES Framework i s implemented as a collection of
Validity: If the information available for a task has a physically and logically distributed nodes communicating
high confidence level, then that task should be per- via an Ethernet LAN. Each node has a similar internal struc-
formed before a task for which the information has a ture consisting of three principal processing components
low confidence level, as the resulting information (see Fig. 34): a blackboard-based expert system, a focusing
would be more valid. tool, and a communications switcher. The expert system
* Internal/External Control: Tasks that are triggered by component exchanges information with the external world
local activity may be preferred over tasks that are via a user interface and a message stream; its own knowl-
received from other experts (and vice versa). edge sources use this information, in conjunction with what

FRANKLIN et al.: EXPERT SYSTEM TECHNOLOGY FOR THE MILITARY 1351

 r M a n u f a c t ' u e of Nuclear Weapon
Triqger by Terrorists

Warninq 2 :

Manufacture of Nuclcar Device by

Terrorists

Users

Expert System
External
I Messages
Fig. 34. PRC's CCES framework

is received from other nodes, t o produce new inferences

and goals. The focusing tool uses a frame-based data struc-
ture describing goal rating factors t o reevaluate items on
the agenda; each node may use different criteria for agenda
evaI uat io n . The commu n ic-atio n switcher exchanges IocaI Iy
generated inferences and goals with other nodes, and
maintains records that allow it t o broadcast local changes
to previously transmitted information.
These three components run as asynchronous, concur-
rent processes in each CCES node (although a single com-
munication switcher can service multiple nodes that are
located in the same physical processor). This type of archi-
t ect u ra l deco m position s ho u l d f ac i I itat e the repIac ement
of local node expert systems with other "foreign" expert
systems or perhaps more conventional ADP components.
This potential adaptiveness is important to the overall goals
Fig. 35. Example of local blackboard
ot PRC's research in distributed problem solving, which
inrlude the need tor a capability to coordinate heteroge-
neous nodes such as expert systems and conventional data
base systems. report items), and an indication of the expert's confidence
In a current test scenario, the CCES Framework is con- in the item. The CCES Framework currently uses an adap-
tigured to support four independent expert nodes, which tation of Dempster-Shafer technique for representing
collectively work on two problems at the same time. The uncertainty (see Section II and Appendix 1 1 ) . The confi-
tour experts are: Blue Intel, which is assigned a current dence is displayed graphically as a horizontal bar: the length
intelligence problem (monitoring the activities of Terror- of the solid left end indicates the degree of belief for the
Ists); Blue Recon, which primarily receives reports of visual item, the length of the solid right end indicates the degree
contact with Red Forces; Blue I&W, which is assigned an of disbelief for the item, and the length of the fuzzy middle
I&W problem (monitoring the status and intentions of Red part indicates the remaining, unaccounted-for belief (or
Forces); and White Intel, which is a source of allied intel- uncertainty).
ligence. Each expert receives a different set of reports about As part of each expert's blackboard display, the CCES
the world, but they all have access to the same knowledge Framework also maintains an agenda display window that
sources. Thus, each expert's name was chosen based on its describesthecurrent statusof that expert'sagenda(see Fig.
data sources and assigned responsibilities, rather than on 35). Unsatisfied knowledge source activations are displayed
any unique local knowledge. in the order of their estimated value (also shown), along
As reports and inferences (called hypotheses) are received with the names of the knowledge sources from which they
or produced by an expert, they are added to the expert's originate. Within the current test scenario, efficiency and
local blackboard. Any knowledge source activations trig- task age are the only criteria used t o evaluate the knowl-
gered by a new blackboard item are then added to the edge source activations.
expert's agenda. Each expert maintains an interactive user The communication policies for this scenario generally
intertace t o these data structures (see Fig. 35). Items are dis- allow one expert t o transmit data to other experts only when
played on the blackboard from top t o bottom in the order it has first received a request for a pattern of information
in which they were created. The visual representation of a that matches that data. For example, the Blue Intel expert
blackboard item includes the nameoftheoriginating node, receives reports about various resources acquired by Ter-
the typeof intormation contained in the item, a brief textual rorists (such as nuclear weapons expertise). These reports
description of the item (with times and grid locations for trigger the activation of knowledge sources capable of

PROCEEDIN(,I OF T H t Ittt, V O L 76 NO 10, OCTOBER 1988

hypothesizing what can be manufactured from the newly F. Cautions
acquired resources (such as a nuclear weapon trigger).
The DPSS paradigm may eventually extend the applica-
However, since the Blue Intel expert's knowledge source
bility of expert systems technology to cover many intrac-
i s unable todetermine locallywhether or not Terrorists have
table military problem domains, but it i s not a panacea. In
acquired other resources required for manufacturing the
addition to spawning new problems (e.g., determining
same product (such as high explosives), it sends a pattern
when a DPSS hasactuallyfinished solvingagiven problem),
description of the missing information as a subgoal to the
it may make existing problems, such as system security,
other experts. Until they receive this request, they will not
more complicated. Nor it is likely that current focus of
transmit this type of information to Blue Intel.
attention mechanisms or communication policies will be
A conservative communication policy of this type is
efficient enough t o prevent a real-time DPSS from over-
appropriate when it is important t o reduce the likelihood
loading itself with inter-expert communications (the DPSS
of experts flooding each other (and the scarce communi-
equivalent of thrashing in a virtual memory architecture).
cation resources) with useless information. However, there
Research at the University of Massachusetts [90], [92] sug-
are some circumstances where it may be desirable for one
gests that DPSS performance may be improved by incor-
expert to transmit data t o other experts, even when they
porating sophisticated nodes that can reason about their
have not expressed an interest in it. For example, when the
own plans (potential task sequences) and the plans of other
Blue Intel expert hypothesizes that Terrorists are manu- nodes. Whether these advanced techniques will be suffi-
facturing a nuclear weapon trigger (and later, a nuclear cient to support fielded operational systems remains to be
device), its local knowledge source recognizes that these
seen.
hypotheses are important enough t o be treated as warn-
Nevertheless, the distributed problem-solving paradigm
ings. Warnings are displayed locally i n a special window
represents a promising direction of research i n expert sys-
(see the top of Fig. 35) and are then transmitted t o all known
tems technology. Without it, someof the most critical, com-
experts, where they are also displayed. plex military problem domainswill remain without the ben-
As additional assistance for human users and experts,
efit of support from expert systems.
each expert node maintains a situation map (see Fig. 36) that
presents a symbolized, geographic representation of items VI. AND FUTURE RESEARCH
CURRENT TOPICS
on the expert's blackboard. When two or more nodes share
the same physical processor, they also share the same sit- This section summarizes selected expert system appli-
uation map. As the blackboard is updated, the situation map cations that have been used for the military. The remainder
i s modified to reflect the changes. Icons o n the map and of the section addresses the key research issues that must
their corresponding entries o n the blackboard can be be solved to enhance the use of expert system technology
selected by the user. Additional information i s available for for future problems.
a selected item, including the tree of knowledge sources,
A. Summary Applications
hypotheses, and reports that were used t o generate the
item. This kind of dependency tracking not only explains The expert system process was described briefly and then
how a given hypothesis is derived; it also supports a net- several key applications were presented. The first appli-
work-wide truth maintenance capability that allows warn- cation was a resource allocation problem that was devel-
ings and hypotheses to be withdrawn if their supporting oped t o help the Marine Corps match weapons or targets.
data i s later retracted. New concepts were described that reduce the amount of
questions that were asked and thus reduced the total
amount of time t o solve the problem.
A second expert system evolved from this first example
thatwas used to help the operator identify complex images.
...................
...................
................... This system will enhance the performance of existing mil-
...................
...................
.I:...................
itary sensors by guiding the operator along a near-optimal
Graat,Wastarn
solution t o the identification of unknown targets. The third
................
................ expert system example examined the area of troubleshoot-
ingelectronic equipment. Prototypes have been developed
that can reason about varying fault conditions and can guide
the operator to identify the faulty component and t o fix the
equipment. The fourth expert system application was a dig-
Golden Plains ital radar imagery exploitation system. This system pro-
duced interpretations of military situations from radar
imagery.
Next, an intelligence analysis system was described that
had a learning component. Machine learning is crucial to
Manufacturing Process Hypothesis avoid the brittleness of expert systems that i s commonly
Military Buildup Hypothesis encountered, and this learning system indicated encour-
Observation Hypothesis aging progress. The final system description was a depar-
Pending Anack Hypothesis ture from conventional isolated expert systems. This dis-
Visual Contact tributed problem solving approach used communicating,
Special Contact
cooperating expert systems that work as a team t o solve
Fig. 36. Situation map. large complex problems.

FRANKLIN et al.: EXPERT SYSTEM TECHNOLOGY FOR THE MILITARY 1353

B. Research Issues problems with the actual language of rules, frames, etc.,
The examples of expert systems in the military presented and facts (tolerances, expected life-time and anticipated
in this paper illustrate that a great deal has been learned temperature and environmental conditions).
about knowledge, its representation, acquisition, and uti- The second major issue in this knowledge acquisition
lization sincethe beginningsofAl research. We believethat bottleneck involves the completeness of the knowledge
there are many reasons to be optimistic about the future base. It is necessary t o determine if this collection of knowl-
applicationsof expert systems t o real-world commercial and edge is enough to solve the actual problem. The complete-
military problems but there are critical research issues t o ness issue can be solved for logic-based knowledge systems
be solved. In a recent presentation Friedland [93] made sev- but no technique has been developed t o solve the general
eral interesting observations o n the future research issues problem. As systems evolve and the knowledge acquisition
in expert systems. He points out that the initial researchers process continues through the life of the systems, as indi-
in AI felt that the power of expert systems would be in com- cated in the first bottleneck, this problem will becomeeven
plex, general-purpose inference engines able to emulate moredifficult. The ultimate solution involves sufficient deep
human beings. Friedland suggests that this led t o a great knowledge that can understand underlying principles in
deal of frustration and the growing realization that the the realms of physics, chemistry, electronics, etc., t o assess
power was in the knowledge. Feigenbaum refers to this as the completeness of i t s knowledge base.
the Knowledge Principle. A very important point i s that vast Large Knowledge Base Manipulation: This bottleneck
amounts of knowledge must be acquired and codified in involves the problems of how data, procedures, and other
the machine. Friedland shows that there are four major bot- problem-solving techniques must be manipulated i n very
tlenecks in the utilization of this knowledge. They are listed large expert systems containing vast amounts of data. Tra-
here and then discussed: ditional techniquesof forward and backward chainingwith
rules, theorem proving, and object oriented programming
knowledge acquisition may not be adequate as the expert systems of today scale
* knowledge consistency and completeness u p t o the large systems of tomorrow.
large knowledge base manipulation The present techniques in knowledge manipulation may
* interface technology. be adequate if there are significant increases in the power
and speed of computer hardware. Undoubtedly, hardware
Knowledge Acquisition: This involves the process of get- improvements will help solve this problem, but another
ting the knowledge into the system. The standard use of a generation of reasoning systems i s needed t o solve these
knowledge engineer i s to act as an intermediary, which can large problems more efficiently.
cause delays and a loss of accuracy. Major research efforts Research in this area involves the direct solution t o large
directed at eliminating this intermediate step have devel- knowledge base manipulation, o'r an attempt at distributed
oped two techniques. The automatic technique would use problem solving. One distributed problem solving solution
inductive and deductive means to acquire the knowledge i s to break the problem into many smaller communicating
without the knowledge engineer. This is clearly related to cooperating expert systems as discussed in Section V. Dif-
certain aspects of the learning projects such as Michalski ficulties t o be solved include the degree of independence
[63]. The semi-automatic technique would use the domain in the individual expert systems, the required amount of
expert as the knowledge engineer, which has become communications among the individual problem solvers, the
increasingly popular in the development of expert systems. degree of autonomous behavior, and the use of robust
The knowledge acquisition process, which can take from learning techniques.
50 to 90 percent of the time and effort in the development Interface Technology: This bottleneck involves the pro-
of an expert system, i s not a single event process. The real- cess of getting the information out of the machine and into
istic view is that the knowledge must change throughout a form that humans can understand and use effectively.
the life of the system. Present systems do not use very much natural language t o
Knowledge Consistency and Completeness: This bottle- explain t o the user what i s happening and why. This prob-
neck involves the combination of knowledge in many dif- lem will become increasingly difficult as the sizes of sys-
ferent forms, from numerous, often disparate, sources with tems increase and the user requires more explicit expla-
a second issue of whether there i s enough knowledge for nations from the expert system.
functional adequacy. As large expert systems are con- Other problems, in addition t o knowledge acquisition,
structed, the sources of information will increase because are:
no single human can possibly know more than a small per- 1) System Brittleness: Today's systems cannot determine
centage of the final system. Another complication i s that if the problem they are trying t o solve is within their "area
much of the information will come from nonhuman of expertise." For example, there is a definite possibility
resources. I n the example i n Section Ill o n the expert sys- that a system designed t o diagnose medical problemscould
tem for maintenance and troubleshooting, much of the data easily be confused between measles and rust o n a car. I n
comes from instruction manuals, design information, com- addition, the systems expect their inputs i n a fairly tightly
puter-aided design programs as well as data from the defined environment and changes in language or syntax
machine designers. The use of all of these different sources can often result in disarray.
can lead to both apparent and real disagreements. The 2) Common Sense Reasoning and Analogical Reasoning:
future research in this area involves the design of mech- A human, stumped by a particular problem, will often try
anisms to aid in the information entry in large systems. t o use"common sense" to help i n the solution. There i s not
Knowledge source disparitycan include contradictory heu- consensus o n what common sense means but Friedland
ristics (likely failure modes and design considerations), [93] suggests that common sense means knowing a little

1354 PROCEEDINGS OF THE IEEE, VOL. 76, NO. IO, OCTOBER 1988
about a lot and recognizing from a human’s vast knowledge 5) Generic Expert Systems: It i s hoped that domain inde-
base what i s relevant and can be linked t o the particular pendent expert systems can be used as financial advisers,
problem at hand. Lenat [94] at Microelectronics and Com- medical diagnosticians, indications and warning analysts,
puter Consortium (MCC) is attempting to codify the knowl- target identifiers, and electronic troubleshooters. No single
edge in adeskencyclopedia,which will be used asthe basis system can accomplish this. Indeed many medical diag-
for the knowledge contained in an expert system. An exam- nostic expert systems cannot be easily modified t o work o n
ple that Lenat uses to show the difficulty of common sense apparently similar medical problems.
reasoning i s the problem of distinguishing the difference Another issue with generic expert systems concerns a
between the t w o phrases: uniform approach to handling uncertainty. The present sys-
tems are ad hoc and usually not mathematically verifiable.
the pen is in the box
When certain information i s concatenated, the problem
the pen is in the corral.
usually becomes more difficult because the propagation of
Lenat hopes that once he has entered the knowledge con- uncertain information is not well understood. Knowledge
tained in his disk encyclopedia, the computer can deter- representation isalso not handled uniformly. Further details
mine the difference in the two uses of pen. Initially Lenat are found in Appendixes I and [I.
felt that learning would provide a solution but now he feels 6) Delivery Machines for Expert Systems: There has been
that hand crafted knowledge and hard w o r k w i l l be the solu- a big change in attitudes in the AI community. Originally
tion for the next 10 years. everything was done o n large high priced LISP machines
The second mechanism i s analogical reasoning. The pro- with identical delivery machines. The user community i s no
cess consists of two steps. The first step is to pick a potential longerwillingto paythis price since conventional hardware
analogous situation and the second step is t o extract the has become increasingly powerful at lower cost. The new
common items. This sounds simple but i s extremely diffi- m icroprocessor-based workstations wil I be the delivery
cult, and little progress has been made in this area of basic machine in manycases,and for some usersthe AI solutions
research. Progress made in the research on common sense will be provided o n a mainframe. The development will still
reasoning may help t o solve this problem too. be on a mix of special AI machines and more general pur-
3) Learning: This area, discussed briefly in Section IV, pose microprocessors.
needs additional development. Clearly there is a large gap 7) The Degree o f Expert in an Expert System: It has
between present-day expert systems and their human become very clear that expert systems are nothing of the
counterparts when it comes to learning. Friedland [93] kind.The majorityof expert systems areonlyslightlyclever.
points out that we would think a fellow human,being hope- This i s not all bad. The trend in expert systems should be
lessly stupid if the same mistake were repeated endlessly. to improve the user from an average grade of poor to medi-
This i s exactly what happens to most expert systems when ocre, not mediocre t o great. The reason i s obvious. Most
they d o not have sufficient knowledge and reasoning com- of the problems that we encounter are routine, time con-
ponents t o solve a problem. Learning becomes mandatory suming, boring and mundane, and it is these irritating little
when one considers that large systems, such as the NASA problems that take from 50 to 80 percent of the human
space station, are continuously evolving. Long-term mis- expert’s time. The solution i s t o use our “clever” expert sys-
sion projects will need learning t o acquire new knowledge tems t o solve the mundane and allow the human to free u p
and to re-validate the system as knowledge i s added. The his time, by as much as 40 t o 50 percent, to solve the prob-
addition of learning to future expert systems will probably lems that the machine cannot solve.
provide the biggest change from today’s expert systems t o These are genuine problems that must be addressed and
those of tomorrow. Potential aid for future learning lies in solved before AI can be used successfully in the complex
the fields of neural networks and causal modeling of phys- world of the present. I n fact, the Defense Advanced
ical systems. Research Project Agency, the Army, the Air Force and the
4) Synergism Among Intelligent Agents: This area has Navy are addressing all of the key issues listed above. The
great potential. The idea, related to communicating coop- success of this research, by some of the best researchers
erating expert systems, or problem solvers, involves using in the United States, i s expected t o hasten the routine use
them t o provide synergism. This can only be accomplished of expert systems bythe militaryand in thecommercial mar-
when the problem solvers are organized as teams able t o ketplace.
work effectively and cooperatively t o solve a bigger prob-
lem. The organization of these problem solvers can be in
APPENDIXI EXPERT SYSTEM TUTORIAL
many forms including hierarchical, committee, distributed
in functional or geographic terms, etc. The final synergistic Expert systems [6]-[IO] are computer programs that
configuration i s ultimately related to the actual problem. attempt t o imitate real experts. The desired final product
One major advantage of these cooperating problem solvers i s computer program outputs with the same correct results
is that individual ones can be made relatively small. The asthoseof human experts.This section touches upon many
state of the art in small expert systems i s far advanced t o of the tools and techniques commonly used in the con-
what can be done for very large expert systems, because struction of expert systems. A general expert system archi-
it is easier to test them, there i s increased confidence, and tecture, knowledge representation schemes, and reason-
production costs are lower. The individual expert systems ing or inference mechanisms aredescribed. Many different
can then be hooked together as a team t o solve a bigger definitions of expert systems exist; however, the objective
problem. The existing examples of this at the University of is a computer program that reasons at least as well as an
Massachusetts, Stanford, and PRC are showing signs of expert in a given field. This i s a lofty goal and in general i s
promise for large synergistic systems. difficult to achieve. A less ambitious, and more successful,

FRANKLIN et al.: EXPERT SYSTEM TECHNOLOGY FOR THE MILITARY 1355

use of expert systems i s not to replace the human expert In words, this example indicates “for all X, Y and Z, and the
but to provide him with expert or knowledgeable assis- case that X i s smaller than Y and Y i s smaller than Z, then
tance. There are many commercial expert systems [95] but this implies that X i s smaller than Z.”
they are used mostly in nonmilitary applications. Some of
these successes include PUFF, an expert system to diagnose Frames, Objects and Semantic Nets
pulmonary disease; NAVEX, an expert system to monitor
controls of space shuttle flights; Cooker Advisor, used t o An alternate form for representing this knowledge is a
troubleshoot electrostatic soup “cookers” for Campbell frame. A frame i s a means of representing a structured sit-
Soup; STEAMER, a system that trains steam plant operators uation such as a typical day in school or a generic object
for Navy ships; ACE, a system to troubleshoot cable; definition. In the frame there are slots that hold different
OCEAN, an expert system that checks orders and config- pieces of information that are important to the particular
ures NCR computers and DECs; and XCON, used to con- stereotyped situation or are features or attributes of the
figure VAX computers. This section will give an overview object. Some of this information concerns how the frame
of expert systems, discuss some of the challenging prob- can be used, or what one can expect to happen or actions
lems, and finally review in more detail the problems asso- that should be taken if certain situations did not take place.
ciated with dealing with uncertain information. Slots are given values to represent instances of a situation
Fig. 1 i s a block diagram of an expert system including or object. For example, a target frame may have slots for
outside interfaces. The knowledge engineer works with a target type, number, manpower requirements, and indi-
domain expert to acquire the critical information that is cations associated with the possible presence of the target.
needed to clone the expert. The process of knowledge Semantic networks are associative networks that link
acquisition i s what the knowledge engineer uses to extract nodes together with lines or arcs. The nodes represent an
data, knowledge, and techniques from the domain expert. object and the arcs joining the nodes represent the rela-
The extracted information includes system facts and sup- tionship between the nodes, such as an engine i s a part of
positions that might have varying degrees of uncertainty a car. In this case the nodes of the semantic net are engine
associated with them. Typically, the knowledge engineer and car, and the arc joining the nodes defines the rela-
will derive from the domain expert the equivalent of a data tionship “is a part of.” Semantic networks are frequently
base of facts that will be put into the knowledge base. This stored in frames.
knowledge base is custom fitted to the domain. It is con-
venient to think of the knowledge base as made up of two Sta te-Space
components: assertions and knowledge relationships. The
State-space was one of the earliest representation for-
assertion component is similar to a working memory or a
malisms and it was developed for problem-solving and
temporary storage of data. It contains declarative knowl-
game-playing programs. The search space is not actually a
edge about a particular problem, and the current status of
knowledge representation but is really a structure of a prob-
the problem that i s being solved. The data in the assertion
lem in terms of the alternatives available at each possible
component of the knowledge base can be represented as
state of a problem. An example would be the alternative
first order predicate logic, frames, semantic networks, state-
moves available on each turn of a game, such as chess. A
space or other techniques that are convenient to that par-
straightforward way of finding the winning move would be
ticular problem. These terms for knowledge representation
to try all of the alternative moves and then try all of the
will be explained in the following paragraphs.
opponent’s responses. Clearly, in complex situations such
as chess there are too many possible combinations and this
Logic
leads to a combinatorial explosion. Special search strate-
A predicate is simply something that asserts a fact about gies must be developed to deal with this combinatorial
one or more entities and has a value of true or false. An explosionorthe run timeof theexpert systemwill beunrea-
example is”Jackcaught the ball.” In thiscaseweare saying sonable. “Knowledge up, search down” i s one way of
that there i s a catching relationship between Jack and the expressing an AI approach to reducing the search space.
ball. Predicate calculus i s a means of calculating the truth The more knowledge that the problem-solving system can
about propositions and it combines the notion of predi- apply to guide the search, the quicker the search will be.
cates with logical relations such as and, or, not, imply and
equivalence. Predicate calculus alone tends to be somewhat Procedures and Algorithms
clumsy to use, so two additional concepts were added. The
Another alternative for knowledge representation i s pro-
first is the idea of operators or functions, and these func-
cedural representation. One of the earliest knowledge rep-
tions are different from predicates because they are not
resentation implementations was on PLANNER [7]. The pro-
restricted to the values of TRUE or FALSE. They can return
cedural knowledge was used to encode explicit control of
objects, and thus a function “uncle o f ” when applied to
a theorem-proving process within a logic-based system. In
Mary would return a value of John. The second additional
this case, the procedural representation contains knowl-
concept i s that of the predicate, equals, which says that two
edge of the world in small programs that know how to do
individuals X and Yare indistinguishable under all predi-
specific things and how to proceed in well specified situ-
cates and functions. With these changes, we have a variety
ations.
of first order logic, and it is no longer pure predicate cal-
The knowledge base alternatives for the selection of the
culus. A simple exampleof first order predicate logic would
knowledge representation schemecould itself usean expert
look like this:
system as an advisor to the human engineer. The length of
v XYZ(SMALLER(X, Y ) SMALLER(Y, Z ) + SMALLER(X, Z). this paper limits the available discussion of this important

1356 PROCEEDINGS OF THE IEEE, VOL. 76, NO. 10, OCTOBER 1988
 topic of knowledge base alternatives. In the specific appli- these smart strategies i s heuristic search, which can usually
cations in Section Ill the reader will be able to see the rich- solve a problem but without a mathematical guarantee.
ness and variety that is required to handle knowledge rep- Generally, heuristic search uses rulesof thumb, tricks, sim-
resentation properly. plifications or any other clever techniques that drastically
The second part of the knowledge base, shown in Fig. 1, limit search in a large problem space.
i s the knowledge relationships component. The most com- Space here limits a full description of these search tech-
mon form that is encountered is the production rule. A typ- niques, but frequently a combination of algorithms and
ical rule is in the form of IFTHESEANTECEDENTSARETRUE heuristics are used. As an example, a particular path on a
THEN THIS CONSEQUENTIS TRUE. Some production rules decision tree could be evaluated by how much cost was
are derived from physical relationships that are defined by consumed to a given point, and then a heuristic would be
conventional science. These rules can often be extracted used to estimate the cost to complete the evaluation of the
from the domain expert or codified information such as a tree from the given point to the final desired goal. These
textbook. Other production rule relationships are derived heuristics are usually determined by interviewing the expert
from empirical forms of knowledge and are generally and determining how he estimates that one branch of a
referred to as heuristics. Heuristics are rules of thumb that decision tree is significantly morecostlyor hasa higher pay-
the domain expert uses, which usually achieve the correct off to evaluate than another. These rules of thumb are then
answer or desired results. These heuristics are not optimal embedded inthe inferenceengine inorderthat itcandecide
in a mathematical sense, but they frequently succeed. what branch of the tree should be evaluated and what
Often, an expert will use hard physical facts combined with branch i s too costly to evaluate or does not have an expected
these rules of thumb to solve his problems successfully. high payoff. The inferenceengine i s then used todetermine
A second type of knowledge i s algorithmic. The knowl- what knowledge should be used from the knowledge base,
edge relationships portion of the knowledge base can con- what inputs should be used, what information should be
tain algorithms. These algorithms are additional procedural transmitted or sent to the user, and how the problem should
knowledge that allows various facts and inputs to be com- be addressed. As information from sensors or other inputs
bined in a calculation that will provide new information in becomes available the inference engine will try and reason
the process of solving that specific problem. Examples of about the situation. The inference engine attempts to guide
the algorithms are estimation techniques to calculate the the expert system to do the required actions to solve a spe-
important parameters such as speed of an object, location, cific problem.
relative motion, etc. The inference engine in its role as agatekeeper can guide
The components of the knowledge base for an expert sys- the expert system toward the solution of the problem. One
tem have been described in a very general manner. The approach called forward chaining reasons forward from the
knowledge base can contain the following: inputs of the expert system toward the final solution. The
object is to guide the present situation forward from its ini-
facts
tial conditions (usually data) to the final situation that will
rules and procedures
satisfy the goal condition. An example would be a chess
logical relationships
game in which you are given a configuration of white and
algorithms
black pieces and you want to use legal chess moves to
heuristics.
achieve checkmate.
One important thing to remember is that the knowledge The second major technique available to the inference
base i s specific to the particular problem that is being engine is to backward chain. Backward chaining involves a
solved. The second major portion of the expert system i s strategy that works from the goal of the expert system and
the inference engine, which will be discussed next. The not from the current situation. Usually the goal statement
inference engine, as opposed to the knowledge base, may is broken into one or more subgoals that can be solved more
be common to a number of domains with similar charac- easilythan the major ultimategoal.The individual subgoals
teristics. The inference engine, shown in Fig. 1, i s a gate- can be broken down further until each sub-subgoal is atriv-
keeper between what the expert system believes and the ial problem that can be solved. This approach i s used when
actual expert system program [q.This gatekeeper is respon- there are a limited number of final states that can be
siblefor adding and deleting beliefs and performing certain achieved or when there i s external evidence to suspect a
classes of inferencing. This inferencing can be done as facts specific situation. For example, for a medical doctor who
are added or when requests for information arrive at the is diagnosing pulmonary disease, he knows that there i s a
gate keeper. relatively small set of diseases that are under consideration.
One way of thinking about an expert system i s to lay the In this case, one expert system called PUFF uses backward
problem out as a graph network that is filled with branches chaining from a specific diagnosis and evaluates the evi-
containing andlor logical functions. In this case, the goal i s dence to determine if this diagnosis i s justified. As the SYS-
to find a node that will solve the problem. The search space tems become more complex, multiple initiative approaches
that must be investigated in the pursuit of the solution can (that combine forward and backward chaining) are fre-
be very large. Barr and Feigenbaum [6] point out that the quently used by the inference engine to solve a given prob-
number of different complete plays for an average length lem.
chess game i s For checkers the search space is esti- Asthe problem domain increases in complexity, thereare
mated to be IO4'. Clearly an exhaustive search of this space definite tradeoffs associated with the level of reasoning. For
would task the fastest supercomputers for an inordinate some problems the physical system is not modeled in avery
amount of time. The search time of this very large, complex accurate way but rather is greatly simplified. In these cases,
space can be reduced by using smart strategies. One of a shallow reasoning system i s frequently used, and in this

FRANKLIN et al.: EXPERT SYSTEM TECHNOLOGY FOR THE MILITARY 1357

 case the knowledge base will contain rules and procedures 4) Dealing with Uncertainty: This i s the situation that we
that do not rigorously reflect the causality of a system. This face most of the time. Because we cannot count o n a radar
can work in an acceptable manner for some problem detection or a sonar detection with a high degree of con-
domains, but frequently a more sophisticated model must fidence, we must determine how the inference engine can
be invoked and a deep reasoning system must be used. An manipulate the facts and propagatethe evidence into a sat-
example could be an electronic diagnostic problem. If the isfactory situation. Since this is such a crucial issue, the dis-
knowledge base only contains rules such as: ”when the cussion in Appendix II i s provided as an introduction t o the
power amplifier levels are low then examine the power sup- problems associated with uncertainty.
ply output voltage” we have an example of a relatively shal-
low reasoning system. An example of a deeper reasoning
system would use more knowledge about the “physics” of
II
APPENDIX METHODS
OF DEALING
WITH UNCERTAINTY

a situation. An example could be an electronic circuit diag- Approaches t o dealing with uncertainty generally take
nostic expert system i n which components such as tran- some variation of numeric characterization. Numeric rep-
sistors are modeled and the system knows that if the base resentations usually take the form of the assignment of a
current increases the collector voltage will drop. point value (as the application of Bayes’ Theorem or max-
At this point we have just touched the tip of the iceberg imum entropy), intervals o n a range(as i n Dempster-Shafer
with regard t o the complexity and variations of the design Theory) or points within an evidence space. A variant o n a
and implementation of expert systems. The main body of numerical approach with foundations in set theory is
this paper i s concerned with giving the reader a better Zadeh‘s Fuzzy Logic, or Fuzzy Set Theory. A discussion of
appreciation for the application of expert systems and adis- these major methods follows.
cussion of some of the interesting problems associated with The sources of uncertainty in reasoning with expert sys-
these applications. We believe that expert systems will play tems are numerous [12]-[15]. Someof these sources include
an important role i n future militaryapplications. Expert sys- situations such as when information i s deficient because
tems research, as well ascurrentlyfielded systems, has illus- it i s partial or not fully reliable, or when the representation
trated that useful systems can be built which displaya great language i s inherently imprecise or information from mul-
deal of intelligence in a narrow domain. Some of the most tiple sources i s conflicting [14]. For example, uncertainty i s
interesting problems t o be solved are briefly listed below. necessarywhen one interviews an expert and receives qual-
1) Acquiring and Structuring a Variety o f Knowledge: itative information such as ”that evidence gives credence
Some of the facets of this problem have been mentioned to this diagnosis or it suggests that this circumstance could
previously. There i s n o clear cut technique, and a hybrid be the cause for this disease.” The problem for the expert
approach has been used i n the past. For example, facts can system designer i s t o decide the acceptable way t o handle
be represented as rules, logic, frames or a combination of this pervasive uncertainty. Chandrasekeran [IS] points out
these techniques and others. that resolution of uncertainty i s something that a human
2) Performing in Real Time: Complex military problems i s expert at doing but that a human does not use a single
with thousands of inputs and hundreds of choices will lit- method for resolving uncertainties. As a result, the expert
erallycausethe machinetogrind t o a halt. Heuristics search system designer needs t o have a bag of tricks t o handle the
and faster computers are parts of the solution, but this rep- different situations requiring uncertainty.
resents one of the concerns for real operational expert sys- For a long time, derivations of the Bayesian model have
tems. been used i n expert systems. The two most well-known rule
3) Dealing with Unanticipated Events:The ability t o learn based expert system examples are MYCIN, an expert sys-
from experience is a key element in intelligent behavior of tem t o diagnose and recommend the therapy for infectious
living organisms, yet one that i s almost totally absent i n blood disease, and PROSPECTOR, an expert system to aid
present intelligent artifacts such as expert systems. The in the identification and location of high valued ore depos-
need for learning in an expert system i s exemplified by the its from geological data. The MYCIN program defined a
vast amounts of knowledge needed in most expert systems cdncept of certainty factors that were used to manage
for good performance. The process of imbuing these sys- uncertainty i n rule based expert systems. I n general, the
tems with knowledge i s slow and error prone and hinges certainty factor concept is an empirical ad hoc technique
upon the developing art of knowledge engineering. The that did not have consistent characterization. Heckerman
prolonged use of expert systems also requires an ability t o [I41 documents a clear, precise and consistent formulation
learn, because as the domain knowledge itself changes, so of the certainty factor model, and the following discussion
mustthe knowledge i n the system, if the system i s to sustain shows how he developed the mathematical foundation for
good performance. The evolution of knowledge i n expert the MYCIN certainty factor. This development was com-
systems to improve system performance must be of par- pleted several years after the expert system was built and
amount importance if the systems are expected to operate tested, and i n general does not bring any clear advantages
effectively in the‘ battlefield of the future. One example of over using a probabilistic model.
learning i s for the machine to consider alternatives with i t s MYCIN’s knowledge is stored as rules in the form of “If
present knowledge base and inference engine and to evidence Then hypothesis.” Frequently in medicine there i s
acquire new techniques. Encouraging projects at Stanford, significant uncertainty between the relationship of evi-
University of Illinois, MIT, University of Michigan and oth- dence and hypothesis. The certainty factor was used t o
ers show that there i s hope but that a great deal of work accommodate these nondeterministic relationships. The
must be accomplished before we have truly autonomous certainty factor varies between -1 and + I . Positive num-
expert systems that can adapt t o a situation and then re- bers convey belief in the hypothesis from the evidence,
structure their rules. whereas negative numbers correspond t o a decrease in

1358 PROCEEDINGS OF THE IEEE, VOL. 76, NO. IO, OCTOBER 1988
belief. These certainty factors d o not correspond t o mea- where CF(H, E') denotes the combined certainty factor
sures of absolute belief. A convenient notation i s shown
below z = } wx w r o

E + CF(H, E ) + H -wy w < o

where H i s the hypothesis, E i s the supporting evidence and w = CF(E, E')

CF(H, E ) i s the certainty factor associated with the rule. It x = CF(H, E )
i s possible for one hypothesis t o be dependent o n more
than one piece of evidence and, further, for that hypothesis y = CF(H, E )
t o serve as evidence for another hypothesis. Each link
z = CF(H, E').
between evidence and hypothesis will have i t s own cer-
tainty factor associated with it. Fig. 37 shows a typical evi- Note that in the above case, the certainty factor, CF(H, E')
is dependent o n CF(H, E ) which is the certainty factor for
Hgiven that the evidence E i s not true. Heckerman [I41 calls
these sequential and parallel combination rules desiderata,
and he shows that this desiderata should be used as the
definitions of certainty factors. He points out that the orig-
inal definitions of certainty factor as defined by Shortliffe
and Buchanan [I51 were inconsistent with the sequential
combinations of certainty factors and caused major incon-
sistencies. Heckerman [I41 shows that if uncertainty i s to
be propagated through an inference network in accor-
LIKELIHOOD RATIO =
P(ElH) dance with the desiderata for combinations, the evidence
ODDS = O(H/E) * O(H) must be conditionally independent given the hypothesis
O(H) = prior odds of hypothesis, H
and its negation. Further, the inference network must have
O(HIE) = updated odds of hypothesis, H given atree structureasopposed t o t h e moregeneral graph struc-
evidence, E
ture. The certainty factors that are used by Heckerman are
Fig. 37. Evidence link. the certainty factors that are defined by the axioms of the
desiderata. I n particular, he finds that monotonic trans-
formations of the likelihood ratio, h(H, E ) , satisfy the de-
dence link. I n this case hypothesis H, i s supported by evi-
siderata where the likelihood ratio is defined as
dence El and E2. As can be seen, evidence E2 i s actually a
hypothesis supported by evidence E4 and E,. h(H, E ) = P(E\H)/P(EIH)
As an example of how this system works let us concen-
trate o n hypothesis H1 that is supported by evidence El and P(€(H)i s the conditional probability of the evidence E
E2 as shown below. One of the first questions t o ask i s how given that the hypothesis H i s true.
to combine this evidence. The simplified situation is shown
below for the parallel combination of evidence: P ( E ( n ) is the conditional probability of the evidence E
given that the hypothesis H is not true.

The odds-likelihood form of Bayes rule i s shown below:

O(HJEe)= X(H, E, e) O(H(e)

where odds of an event x i s O(X) = P(x)/(l - P(x)) and P(x)

where CF(H1,E,) i s the certaintyfactor of hypothesis H1asso- i s probability of x, and O(H(Ee)i s the odds of hypothesis H
ciated with evidence El and CF(H,, E2) i s defined similarly. conditioned o n evidence E and prior evidence e. Notice in
Theequivalenceof this situation i s shown o n the right-hand this expression that the likelihood ratio provides an update
side but i n this case the evidence El and E2 are shown as a in the odds of a given hypothesis since h i s a multiplier fac-
single input with a composite certainty factor CF(H1,El, E2). tor times the prior odds and this yields updated odds. The
Heckerman points out that the certainty factors are com- likelihood ratio h can vary from zero t o infinity as opposed
bined as indicated below: to the desired value of -1 t o +Ifor the certainty factors.
This problem is easily resolved by setting CF(H, E ) = F(X(H,
x+y-xy x,yso
E))whereFisafunction that mapshintothe interval of [ - I ,

z = [
1 -

x+y+xy
X+Y
min (1x1, lyl)
x, y are of opposite sign

X r Y < O
1 +I]. One function for F(x) is

F,(x) = (X - I)/x, x 2 1

= x - I , x < 1.
where x = CF(H1,El), y = CF(H1,E2) and z = CF(Hl, €,, E d .
One certainty factor CF1(H,E)that satisfies the desiderata
The second situation is concerned with the combination
of evidence that is configured in the sequential configu- i s found t o be
ration shown below: CF,(H, E ) = F1(X(H, E ) = ( h ( H , E ) - I)/X(H, E ) , for h L 1
CF(E €') CF(H E ) CF(H,€7
E " € d H or €'- H = h(H, E ) - 1, for h < 1.

FRANKLIN et al.: EXPERT SYSTEM TECHNOLOGY FOR THE MILITARY 1359

This certainty factor CF, can also be expressed as of an infecting organism, a smear showing gram negative
organisms narrows the hypothesis set of all possible orga-
nisms to a specific subset. A Bayesian approach might
assume equal prior probability and distribute the weight
of this evidence uniformly. Shafer points out that this does
not distinguish between uncertainty (or lack of knowledge)
and equal certainty. I n effect the Dempster-Shafer theory
It is interesting to go back and compare this form with the attributes belief to subsets as well as to individual elements
original MYCIN work. The difference i s an added factor in of the hypothesis set. Zadeh [I81 gives a simple example in
the denominator for each of the above expressions for the explaining the Dempster-Shafer theory. Assume that
certainty factor. CountryX believes that a submarine 5, belonging to Coun-
There are numerous monotonic transformations of the try Y, is hiding i n X’s territorial waters. The Ministry of
likelihood function that allows a probabilistic interpreta- DefenseofXsummons hisexpertSE,, . ,E,,andaskseach
tion for certainty factors and, in fact, Heckerman shows that one to indicate the possible locations of S. The first m
every probabilistic interpretation for certainty factors i s a experts, E,, . . . , E,, m In, give the possible locations as
transformation of A. INTERNIST, an expert system for diag- L,, . . . , L,. Each location L, for i = 1to m i s a subset of the
nosing internal medicine, uses an ad hoc scoring system ,,
territorial waters. The rest of the experts E, + . . ,E, assert
that was very similar to the additive property of the log- that there is no submarine in the territorial waters and the
likelihood system which states that equivalent statement i s that L,, = 0, . . * , L, = 0where
0 i s the empty set.
If the Minister of Defense raises the question, “Is S in a
The GLASGOW DYSPEPSIA system uses a certainty factor specified subset, A, of the territorial waters?”, there are two
equal to In X[(h, e)] whereas PROSPECTOR uses X(H, E ) and cases that arise.
in particular a combination function of Case 1) E, i s a member of A and this implies that Expert
X(H, E ) - 1 E, feels certain that S i s in A.
Case 2) E, fl A # 0 and this implies that it i s possible or
X(H, E )+ 1‘
Dlausible that S i s in A.
There are many problems concerned with the certainty
Clearly case 1) implies case 2).
factor discussed above. First, inherent in the certainty fac-
Assume the Minister of Defense aggregates his experts’
tor model, there must be conditional independence of evi-
opinions by averaging. If k out of n experts vote for case
dence given Hand given not U. Another problem is that the
I), the average certainty (or necessity) i s k/n and if q experts
model does not hold for nontree networks. A third problem
vote for case 2), the average possibility i s q h . If the opinion
for systems such as MYCIN i s that certainty factors were not
of those experts who believe there i s no submarine in the
derived from a clear operational definition. Despite all of
territorial waters is disregarded, the average certainty will
these problems, MYCIN worked as well as an expert, and
be k/m and the average possibility or plausibility i s q/m,
Heckerman pointsoutthat asensitivityanalysis of MYCIN‘s
respectively. The disregarding of those experts that said
knowledge base showed that the system performance did
there is no submarine in the territorial waters is referred to
not change significantly when large numbers of the cer-
as normalization. Zadeh [I81 points out that normalization
tainty factors were changed. It i s important to note that the
can lead to counterintuitive results since it suppresses an
original MYCIN certainty factors made some sense due to
important aspect of the experts‘ opinions.
the application area of medical diagnosis. They tend not to
The Dempster-Shafer theory shows how the weighted
make sense outside that domain-a failing of some expert
experts’ opinions can be combined. For example, if expert
system shells that use the original factors as inherent tools.
E, has an opinion weighted by W,, the average normalized
Kyburg, in a recent AAA1 workshop, notes that under very
certainty P, i s
weak assumptions, probability i s the onlytheorywith point
values that makes sense [96]. Levitt addresses the issues in
”Model-Based Probabilistic Situation Inference in Hierar-
chial Hypothesis Spaces“ [15].
P, =
Kc
-
)
Wl for E, a member of area subset A

and average normalized possibility f pi s

Dempster-Shafer
Alternate approaches t o certainty factors have been
P,, = -
:(, W
1)
for E, that could be in subset area A

investigated since MYCIN and used on expert systems. One where K = 1 - C W, for E, not a member of subset area A.
alternate theory was developed by Dempster [I61 and an J
application was developed by Shafer [ I q . One advantage The average normalized certainty i s the belief function of
of the Dempster-Shafer theory i s to model the narrowing Dempster-Shafer, the average normalized possibility func-
of the hypothesis set with the accumulation of evidence. tion i s the plausibility function of Dempster-Shafer, and the
This i s a process that characterizes diagnostic reasoning in weights w,, . . , w, are basic probability functions of
medicine, troubleshooting of electronic equipment and Dempster-Shafer’s theory.
expert reasoning in general. Frequently an expert will use If the Minister of Defense wanted to know what the prob-
evidence that will focus his thinking on a larger subset of ability f ( A )was that S i s in A , the normalized answer would
the total possibilities as opposed to a single hypothesis. For be
example, Shortliffe [I51 points out that in the identification f , IP(A) 5 Pp.

1360 PROCEEDINGS OF THE IEEE, VOL. 76, NO. 10, OCTOBER 1988
 The basic components of the Dempster-Shafer theory I n the same paper Zadeh presents what i s probably one
are the representation and the rule for combining evidence of the more implementable views of Dempster-Shafer the-
or degrees of belief i n evidence. A major distinction i s that ory. He views it as applied to relational data base tech-
certainty i s not merely allocated t o singleton hypotheses, nology, as an instance of inference from second-order rela-
but also t o sets of hypotheses; this leaves some belief per- tions. Zadeh relates the measures of belief and plausibility
haps unallocated t o a definite hypothesis, but allocated tothecertainty(or necessity)and possibilityof agiven query
nonethelesstoasetwhich maycontain thecorrect hypoth- set Qfor retrieval from a second-order relation in which the
esis. The allocation of belief i s in the construction of belief data entries are possibility distribution.
functions over a set of hypotheses 8. (This set of hypotheses, As an example, Zadeh shows a database called EMP2 that
or exclusiveand exhaustive possibilities, iscalled the frame contains the following information:
of discernment.) These belief functions are mappings from
EMP2 DATABASE
the power set of 8 to the unit interval, such that the belief
in the null set is 0, and the belief in 0 is 1. NAME AGE
In the framework of an I&W system, 8 could be the set
of observable Soviet missiles, with a subset of strategic 1 [22, 261
offensive missiles and the growing family of Soviet Mobile 2 [20, 221
Missileswithin the subset. The classification could proceed
down t o the singleton sets, containing such missiles as the 3 [30, 351
SS-16, the SS-20, the SS-25, the SS-X-24, etc. The subsets of 4 [20, 221
0 f o r m a k i n d o f membershiptree,with Qatthetopextend-
ing down to the singleton sets at the bottom. The Dempster- 5 [28, 301.
Shafer basic probabilityassignment (bpa),massignsaquan- Thus, in the case of category name, 1, the interval value for
tity of belief t o every element i n the tree. This bpa corre- the age of 1 is known to be in the set (22, 23, 24, 25, 26).
sponds to the weight w that was discussed before. A belief This set contains the possible values of the variable AGE (1)
function Bel, (which corresponds t o the average normal- or equivalently i s the possibility distribution of AGE (1). A
ized certainty discussed earlier) represents the belief i n a query Q t o this database can ask a question such as “what
subset; Bel entails the belief in all subsets contained in that fraction of employees satisfy the condition that AGE(;) E Q,
set by combining the values for m(A), for all subsets A. The i =I;* * ,5,whereQisthequeryset[20,25].”Zadehpoints
belief function o n a subset of mobile missiles, the 5 1 6 , the out that the query set Q and the data entries in the column
SS-X-24, and the SS-25, would be represented by the sum labeled AGE can be regarded as a possibility distribution.
of the basic probability assignments o n all the subsets of In this context, Zadeh says the database information and
that subset; Bel({SS-16 SS-X-24 SS-25)) = m({SS-16 SS-X-24 the queries can be described as granular with the data and
+
SS-25)) m({SS-16 SS-X-24)) + m({SS-16 SS-25)) + m({SS- the queries play the role of granules.
X-24 SS-25)) + m({SS-16)) + m({SS-X-24)) + m({SS-25)). Forthe situations where thedatabase attributevalues are
Uncommitted belief i s belief that i s committed t o the entire not known with certainty, it is proper t o consider the pos-
I
set 8; m(O). Total ignorance i s represented by the vacuous sibility of Q given the possibility distribution. An example
belief function; where m(O) = 1, and m(A) = 0 for all sub- would be if the query Q was [20, 251 and AGE(1) i s [22, 261,
setsAof8.The belief interval ofasubsetAisgiven by[Bel(A) it i s possible that AGE(1) E Q, it i s certain (or necessary) that
1-Bel(A‘)].Thewidth of this interval i s the uncertaintyof our AGE(4) E Q, and it is not possible that AGE(5) E Q. In general
belief in the hypotheses contained in A. For example, we form we have:
might associate a belief interval [.5,2] with an indicator o n
the above subset; this means that we are 50 percent sure a) AGE(;) E Q i s possible if D, f l Q # 0 = empty set
that the observed target is in the above set, and 20 percent where D, is the possibility distribution.
sure that it i s not. We might use Dempster’s rule of com- b) Q is certain (or necessary) if Q 3 D,.
bination when we get a new belief interval from another c) Q i s not possible if D, n Q = 0 = empty set.
knowledge source. The scheme for combining will update Updating the above table with test results for Q of [20,25],
both our belief and our disbelief in the proposition. The we have:
Dempster rule of recombination can be a problem because
NAME AGE TEST
when sources of evidence are combined, it i s assumed that
they are independent and frequently this is not the case. 1 [22, 261 possible
Zadeh [13], [I81 discusses a serious problem with the
2 [20, 221 certain
Dempster-Shafer theory; he shows that the use of the
method tocombineevidencefromdistinct sources may lead 3 [30, 351 not possible
to counterintuitive results. To modify his example for our
4 [20, 221 certain
military application, suppose sensor A reports that the
observed target is an SS-16 with certainty 0.99, or an SS-20 5 [28, 301 not possible.
with certainty 0.01, but supports the assertion that the tar-
get is an SS-X-24 with certainty 0.99. Applying the orthog- At this point, we can form an answer to the query “What
onal sum of Dempster-Shafer provides us with the unlikely fraction of the employees are in the range of ages equal to
conclusion that the belief i n the assertion that the target is 20,21,22,23,24, and 252”The response, Resp(Q), will be
in two parts, one relating to certainty or necessity N(Q) and
an SS-20 i s 1.0.This i s clearly a problem, and stems primarily
from the fact that under Dempster-Shafer null values are theother relatingto its possibilityK(Q)and i t w i l l bewritten:
not counted, but rather attributed to ignorance. Resp(Q) = N Q ) ; a(Q)).

FRANKLIN et al.: EXPERT SYSTEM TECHNOLOGY FOR THE MILITARY 1361

For our example we have provides some advantages.There will be a great deal of work
necessary before it can be used in a broad group of prob-
Resp[20, 251 = (N[20, 251) = 215; 7420, 251) = 315).
lems as some of the implementation problems have dis-
In this case certainty also counts as possible since certainty cussed. Evenwith the current restrictions, Dempster-Shafer
implies possibility. The first entry for Resp(Q) i s N(Q) and has been used t o help in thedevelopment of expert systems
i s referred t o as a measure of belief in the Dempster-Shafer that deal with uncertainty.
thebry and the second entry a(Q) is the measure of plau- As was stated in the introduction to this section, most of
sibility in Dempster-Shafer. If EMP2 database i s a relation the AI methods for dealing with uncertainty do not attempt
in which the values of age are singletons chosen from the to mirror the way humans process uncertain or inexact
possibility distributions in EMP2, then the response to Q information. Lofti Zadeh [19]-[21] has fathered a branch of
of N(Q) and P(Q) are the lower and upper bounds. AI called Fuzzy Logic, which aims at solving the lack of
Zadeh goes on to show that N(Q) and P(Q) can be com- expressivenessthat plagues classical probabilitytheory. The
puted from a summaryof EMP2which specifies the fraction problem i s the fuzzy language that humans use to describe
of employees whose ages fall in the interval-valued entries the events that we deal with and our beliefs that they will
intheAGEcolumn.AssumeEMP2 hasn rows,with theentry occur, and in particular with the kinds of uncertainty that
in row i = 1, . . . , n under age 0, and that the 0, are com- one deals with in expert systems.
prised of k distinct sets A I , * . . , Ak so that each 0 i s one In probability theory, we cannot represent propositions
of the A,, s = 1, . * * , k. For our example, we have: containing fuzzy language, such as
n= 5 k = 4 fuzzy predicates; tall, old, irascible
fuzzy quantifiers; most, several, few, usually
Dl = [22, 261 AI = [22, 261
* fuzzy events; the boss will be in a difficult mood, this
0 2 [20, 221 A2 = [20, 221 article will gain wide public acceptance, he will be in
good health throughout h i s golden years
0 3 =z [30, 351 A3 = [30, 351 fuzzy facts; older people look more distinguished
0 4 [20, 221 A4 = [28, 301. fuzzy probabilities; likely, not too likely, pretty unlikely
fuzzy rules; i f you want to succeed in business, you've
0 5 = [28, 301 got to know a lot o f people.
If EMP2 i s viewed as a parent relation, its summary can be Fuzzy language surrounds us. It i s intuitive to our rea-
expressed as a granular distribution A of the form soning and thought processes. It is, quite simply, the way
A = { ( A i , PI), (A2t pJt . . . (Akr pk)}
I
we think. Things are not always black or white, and not
always black, white, or gray. Sometimes things are charcoal
in this case p,, s = 1 , . . . , k, is the fraction of the 0's that gray, pearl gray, graphite gray, or soot gray. In order to more
are A,. For our case closely approximate this cognitive representation, Zadeh
A = {([22, 261, 1/5), ([20, 221, 2/5), ([30, 351, 1/5), developed fuzzy logic, building upon traditional set theory
and aiming at the second main timitation with probability
(W, 301, I n ) } . theory: the foundation of two-valued logic on which it i s
In summary, wecan express N(Q)and *(Q) defined in terms built [15]. In classical probabilitytheory, thedescriptive lan-
of the granular distribution 0 as guage i s black or white; an event happens or it does not,
an object is in a set or it i s not, an item has a characteristic
Belief = N(Q) = C p,
5
or it does not. In fuzzy sets, an object i s not simply in or
out of a given set, an object is assigned a grade or degree
such that (Q 3 A,, s = 1, . . . , k) of membership, expressed along the unit interval (0, I),
where 0 stands for non-membership and 1 stands for (def-
Possibility = *(Q) = Cp,
5 inite) membership.
Consider the following basic example for fuzzy set rep-
such that (A, fl Q # 0,
s = 1, . . . , k). resentation:
Note that this interpretation i s consistent with the earlier Let B be a bag of potatoes. What i s the likelihood that the
discussion of Dempster-Shafer and that the p, correspond potato that you pull out to bake i s a huge one? If there are
to the weights wand are the basic probability functions of n potatoes, pl, p2,. * . , pn, you may define the fuzzy set
Dempster-S hafer. HUGE, and the membership function ~HucE(P,),i = 1,
It can be shown that P(Q) = 1 - N(Q') where Q' is the . . . , n which denotes the degree of hugeness of each
complement of Q. In this explanation of Dempster-Shafer, potato. So a 14 ounce potatop, might have pHuCE(px) = 0.78,
Zadeh shows insight as to why normalization causes coun- whereas a 3 oz potato pymight have pHUCE(py) = 0.2.
terintuitive results. He also points out that in the case of Zadeh has defined a way of expressing the number of
definite attributes, the Dempster-Shafer rule for combi- huge potatoes in the bag B, by using a concept called the
nation of evidence is not applicable unless the underlying sigma-count, which i s the sum of the grades of member-
granular distributions are combinable; that is, they have at ship, rounded to the next integer, if appropriate. Further,
least one parent relation which i s conflict free. This implies if the bag has a lot of little potatoes, and one does not want
that distinct probability distributions are not combinable these small potatoes t o misrepresent the HUGE count by
and hence Dempster-Shafer i s not applicable to such dis- providing enough small membership grades to raise the
tributions. HUGE count by an integral number, it i s possible to specify
Despite these problems, the Dempster-Shafer theory a minimum threshold, under which the potatoes will not

1362 PROCEEDINGS OF THE IEEE, VOL. 76, NO. IO, OCTOBER 1988
be counted. As an example, suppose the bag of potatoes ligence to Battlefield Information Management, Battelle
contained 10 potatoes, and we wanted t o know how many Columbus Laboratories, (Washington, DC), Apr. 20-22,1983.
R. Shumaker and J. E. Franklin, “Artificial intelligence in mil-
are huge potatoes. Let us set the threshold at 0.3, and all
itary applications,” Signal Magazine, vol. 40, p.29, June1986.
potatoes with pHUCE under 0.3 will not be considered i n our A. Barr and E. A. Feigenbaum, The Handbook of Artificial
sigma-count. If the membership values arepHuGE(p,)= (0.78, Intelligence. Los Altos, CA: William Kaufmann, Inc., 1981.
0.2,0.5,0.43, 0.25, 0.64, 0.27, 0.3,0.19,0.8), with our thresh- E. Charniak and D. McDermott, “Introduction to artificial
old, the sigma-count would be 3.15, or simply 3 huge pota- intelligence.” Reading, MA: Addison-Wesley, 1984.
P. Harmon and D. King, Expert Systems. New York, NY:
toes. Without the threshold, the sigma-count would be 5.08, Wiley, 1985.
or 5 huge potatoes, which seems t o widen the definition of P. H. Winston, Artificial Intelligence, Second Edition. Read-
huge t o include medium-sized potatoes. ing, MA: Addison-Wesley, 1984.
Zadeh has built upon Fuzzy Logic t o develop Possibility P. Jackson, Introduction to Expert Systems. Reading, MA:
Addison-Wesley, 1986.
Theory, with the same kind of constructs that are found in
D. B. Lenat, A. Clarkson, and G. Kiremidjian, “An expert sys-
probability theory. I n the previous example, one would tem for indications and warning analysis,” IJCAIProceedings,
develop the possibility distribution, so that you could cal- pp. 259-262, 1983.
culate the possibility of choosing a huge potato to bake. R. 0. Duda, P. E. Hart, and N. J.Nilsson, “Subjective Bayesian
This distribution may be similar t o probability distribution; methods for rule-based inference systems,” Tech. Rep. 124,
Stanford Research Institute, Palo Alto, CA, Jan. 1976.
the difference lies more i n the interpretation of the dis- L. A. Zadeh, “A simple view of the Dempster-Shafer theory
tributed values. Let us go back t o our earlier example: of evidence and its implication for the role of combination,”
Let X be a hostile relocatable missile in the set A = { S S - Al Magazine, vol. 7, no. 2, Summer 1986.
20, SS-16, SS-25, SS-x-24). As a result of observances, we D. Heckerman, “Probabilistic interpretation for MYCIN’scer-
tainty factors,” in Uncertainty in Artificial Intelligence, L. N.
could define a probability distribution Pr(X EA) = {0.7,0.1,
Kanal and j. F. Lemmer, Eds. New York, NY: North Holland,
0, 0.2). That is, X is quite probably the SS-20, and highly 1986.
unlikely that it i s the SS-25, as that would be a blatant vio- B. Chandrasekaran and M. C. Tanner, “Uncertainty handling
lation of the SALT treaty. In contrast, also based o n observ- in expert systems: uniform vs. task-specific formalisms,” in
ances, we could describe a possibility distribution, P(X EA) Uncertaintyin Artificiallntelligence, L. N. Kanal and J. F. Lem-
mer, Eds. New York, NY: North Holland, 1986, pp. 35-46.
= { 1, 1,0.7,1}, meaning that it could possibly be any one A. P. Dempster, “A generalization of Bayesian inference,” J.
of the four, but less possible that X i s an SS-25. In possibility Roy. Statis. Soc. Ser. B 30, 1968.
theory, every member of the set could have a possibilityof G. Shafer, A Mathematical Theory of Evidence. Princeton,
one. An obvious difference between a possibility and a NJ: Princeton University Press, 1976.
L. A. Zadeh, “Review of Shafer‘s book, a mathematical theory
probability distribution i s the requirement that the prob-
of evidence,” A/ Magazine, vol. 5, no. 3, Fall 1984.
abilities sum t o one, although possibility values could be -, A ComputationalTheory of Dispositions. Berkeley, CA:
interpreted as maximum probability values. University of California, 1984.
Zadeh claims that possibility theory subsumes classical -, “A theory of commonsense knowledge,” in Aspects of
probability theory. His claim is that possibility theory, and Vagueness, Skala, Termini, and Trillas, Eds. Dordrecht, Hol-
land: D. Reidel Publishing Co., 1984.
its axioms and functions, answers well every problem stated -, ”Syllogistic reasoning as a basis for combination of evi-
for classical probability, and in addition, allows for the rep- dence in expert systems,” in Proceedings of the 9th lnter-
resentation of fuzzy logic and inference from fuzzy sets. In national Joint Conf. on Artificial Intelligence, (Los Angeles,
spite of this apparentlywide potential, fuzzy logic and pos- CA), Aug. 1985.
R. Cohen and M. R. Grinbert, “Atheoryof heuristic reasoning
sibilitytheory have not been widely implemented in expert
about uncertainty,” AI Magazine, vol. 4, no. 2, 1983.
systems. -, “A framework for heuristic reasoning about uncer-
tainty,” in Proceedings of the 8th lnternationalloint Conf. on
Artificial Intelligence, (Karlsruhe, West Germany), 1983.
ACKNOWLEDGMENT R. Cohen and M. D. Lieberman,“A report on FOLIO:An expert
The authors wish to thank Mrs. Ann Goddard for the tire- assistant for portfolio managers,” in Proceedings of the 8th
Internationalloint Conf. on Artificial Intelligence, (Karlsruhe,
less and careful preparation of this manuscript and Mrs. West Germany), 1983.
BetteViolettefor her patient editing of this manuscript. We Cohen et al., “Representativeness and uncertainty in clas-
also thank Dr. Y. T. Chien and Ms. Michelle Younger for sification systems,” AI Magazine, vol. 6, no. 3, 1985.
providing a great deal of guidance and direction. The IEEE P. R. Cohen, “Heuristic reasoning about uncertainty: An Arti-
ficial lntelligenceapproach,” Great Britain: PitmanAdvanced
reviewers’ comments were thoughtful and well received in Publishing Program, 1985.
this revised paper, and we thank them for their useful J. Doyle, “Methodological simplicity in expert system con-
inputs. struction: the case of judgements and reasoned assump-
tions,“ A/ Magazine, vol. 4, no. 2, 1983.
-,“A truth maintenance system,” Artificiallntelligence, vol.
REFERENCES 12, 1979.
M. L. Ginsberg, ”Implementing probabilistic reasoning,”
A. J. Baciocco, “Artificial intelligence,” Military Science Tech- Working Paper HPP 84-31, Stanford University, Stanford, CA,
nology, vol. 1, no. 5, pp. 38-40, 1981. June 1984.
-,“Artificial intelligence and C31,” SignalMagazine, AFCEA, J. A. Barnett, ”Computational methods for a mathematical
Sept. 1981. theory of evidence,” in Proc. of the 7th lnt. Joint Conf. on
E. W. Martin, “Artificial intelligence and robotics for military Artificial Intelligence, (Vancouver, British Columbia, Can-
systems,” in Proc. of The Army Conf. on Application of Arti- ada), 1981.
ficial Intelligence to Battlefield lnformation Management, T. D. Garvey, J. D. Lowrance, and M. A. Fischer, “An inference
Battelle Columbus Laboratories, (Washington, DC), Apr. 20- technique for integrating knowledge from disparate
22, 1983. sources,” in Proceedings of the 7th lnternational Joint Conf.
A. B. Salisbury, “Opening remarks on artificial intelligence,” on Artificial Intelligence (Vancouver, British Columbia, Can-
in Proc. of The Army Conf. on AppIication of Artificial lntel- ada), 1981.

FRANKLIN et al.: EXPERT SYSTEM TECHNOLOGY FOR THE MILITARY 1363

[32] J. D. Lowrance and T. D. Garvey, ”Evidential reasoning: and -,”Utility-based control for computer vision,” in Proc.AAA/
implementation for multisensor integration,” SRI Technical Uncertainty in Artificial lntelligence Workshop, (Minneapo-
Note 307, Dec. 1983. lis, MN), Aug. 1988.
1331 J. R. Slagle and H. Hamburger, ”An expert system for a J. H. Holland et al., Induction: Processes o f Inference, Learn-
resource allocation problem,” Communications of the ACM, ing, and Discovery. Cambridge, MA: MIT Press, 1986, ch. 1,
vol. 28, no. 9, pp. 994-1004, Sept. 1985. pp. 1-28.
[34] J. R. Slagle, M. W. Gaynor, and E. J. Halpern, ”An intelligent A. E. Feigenbaum, “Lecture at the first U.5-China joint sem-
control strategy for computer consultation,” /E€€Trans. Pat- inar on automation and intelligent systems,” Beijing, China,
tern Anal. Machine lntell., vol. PAMI-6, Mar. 1984. May 28-June 1,1984.
[35] L. B. Booker, “An Artificial Intelligence (AI) approach to ship A. L. Samuel, ”Some studies in machine learning using the
classification,” in 20th Annual Technical Symp., Washington, game of checkers,” IBMJ. Res. Develop., vol. 3, pp. 211-232,
D.C. Chapter of ACM (Gaithersburg, MD), 1985. 1959.
[36] R. R. Cantone, F. J. Piptone, W. B. Lander, and M. P. Marrone, M. Minsky, “Steps toward artificial intelligence,” in Com-
“Model-based probabilistic reasoning for electronics trou- puters a n d Thought, A. E. Feigenbaum and J. Feldman, Eds.
bleshooting,” in lJCAl-83. New York, NY: McCraw-Hill, 1963, pp. 429-435.
1371 R. 0. Duda et al., ”Development of a computer-based con- R. S. Michalski, “Understanding the natureof learning: issues
sultant for mineral exploration,“ Annual Report, SRI Projects and research directions,” in Machine Learning, vol. 2, R. S.
5821 and 6415, SRI International, Menlo Park, CA, Oct. 1977. Michalski, J.G . Carbonell, andT. G . Mitchell, Eds. Los Altos,
[38] R. R. Cantone, W. B. Lander, M. P. Marrone, and M. W. Cay- CA: Morgan Kaufman, 1986, ch. 1, pp. 3-18.
nor, “IN-ATE”: Fault diagnosis as expert system guided J. R. Quinlan, “Learning efficient classification procedures
search,” in Computer Expert Systems, L. Bok, and M. J. and their application to chess end games,” in MachineLearn-
Coombs, Eds. New York, NY: Springer-Verlag, 1986. ing, vol. 1, R. S. Michalski, J. C. Carbonell, and T. G . Mitchell,
[39] F. J. Pipitone, ”An expert system for electronics trouble- Eds. Palo Alto, CA: Tioga Publishing, 1983, ch. 15, pp. 463-
shooting based on function and connectivity,“ in / E € € 7st 482.
Conf. on A/ Applications, (Denver, CO), 1984. (651 D. B. Lenat, “The role of heuristics in learning by discovery:
[40] -, ”FE, an electronics fault isolation system basedon qual- three case studies,” in Machine Learning, vol. 1, R. S . Mich-
itative causal modeling,” in Aerospace Applications of A/ alski, J. C . Carbonell, and T. G . Mitchell, Eds. Palo Alto, CA:
Conf., 1985. Tioga Publishing, 1983, ch. 15, pp. 249-263.
[41] -, “The FIS electronics troubleshooting system,” in Com- G . E. Hinton et al., “Boltzmann machines: Constraint satis-
puter, July 1986. faction networksthat learn,”Tech. Rep. CMU-CS-84-119,Car-
[42] J. DeKleer, ”Reasoningwith uncertainty in physical systems,” negie-Mellon Univ., Department of Computer Science, 1984.
Allournal, 1985. J. H. Holland, Adaptation in Natural and Artificial Systems.
[43] R. Davis et al., “Diagnosis based on description of structure Ann Arbor, MI: Univ. of Michigan Press, 1975, ch. 6, pp. 89-
and function,” AAA/ Proceedings, pp. 137-142, 1982. 120.
[44] T. Levitt etal., “Design of a probabilistic certainty calculus for D. E. Goldberg,”Computer aidedgas pipelineoperation using
ADRIES,” Contract DACA76-86-C-0010,Advanced Decision geneticalgorithms and rule learning,” Ph.D. Dissertation,Ann
Systems, Mountain View, CA, Apr. 1986. Arbor, MI, Univ. of Michigan Press, 1983.
[45] T. Levitt, L. Winter, T. Eppel, T. Irons, and C. Neveu, “Terrain R. J. Brachman and 1. G . Schmolze, ”An overview of the KL-
knowledge elicitation for ADRIES: Part 11,” Contract DACA76- ONE knowledge representation system,” Cognitive Science,
86-C-0010, Advanced Decision Systems, Mountain View, CA, vol. 9, pp. 171-216, 1985.
Oct. 1987. J. H. Holland, ”Properties of the bucket brigade algorithm,”
[46] T. Levitt, W. Edwards, and L. Winter, ”Elicitation of a priori in Proc. of an lnt’l. Conf. on Genetic Algorithms and Their
terrain knowledge for ADRIES,” Contract DACA76-86-C-0010, Applications, pp. 1-7, 1985.
Advanced Decision Systems, Mountain View, CA, Nov. 1986. C. L. Forgy, “Rete: A fast algorithm for the many patternlmany
[471 DARPA, ETL, ADS, SAIC, MRJ, and TASC, “Advanced digital object pattern match problem,” Artificial Intelligence, vol. 19,
radar imagery exploitation system (ADRIES) program plan,” pp. 17-37, 1982.
Advanced Decision Systems (ADS), Mountain View, CA, Oct. K. S. Keller and H. J. Antonisse, “Prediction-based compet-
1986. itive learning in the M2 system,” to appear in Proc. ExpertSys-
[48] R. Drazovich et al., “Advanced digital radar imagery exploi- tems in Gov’t. Conf., 1987.
tation system (ADRIES) annual technical report,” Contract H. J. Antonisse and K. S. Keller, ”Genetic operators for high-
DAEA18-83-C-0026,Advanced Decision Systems, TR-1040-02, level knowledge representations,” in Proc. of2nd lnt’l. Conf.
Apr. 1985. on Genetic Algorithms and Their Applications, pp. 69-76,
[49] D. Granrath, “Estimation procedure for ROC curves,” Sci- 1987.
ence Applications International Corporation (SAIC) Internal R. L. Riolo, “Bucket brigade performance: I. Long sequences
Memo dated April 15, 1987, Tucson, AZ, Apr. 1987. of classifiers,” in Proc. of 2nd lnt’l. Conf. on Genetic Algo-
[50] D. von Winterfeldt and W. Edwards, Decision Analysis and rithms a n d their Applications, pp. 184-195, 1987.
Behavioral Research. Cambridge, MA: Cambridge Univer- D. Klahr et al., Production System Models of Learning a n d
sity Press, 1986. Development, D. Klahr, P. Langley, R. Neches, Eds. Cam-
[51] J. Pearl, “Fusion, propagation, and structuring in belief net- bridge, MA: MIT Press, 1987, ch. 1, pp. 1-53.
works,” Artificial Intelligence, vol. 29, pp. 241-288, 1986. J. E. Laird etal., “Chunking in Soar: The anatomy of a general
[52] R. A. Howard and J. E. Matheson, “Influence diagrams,” SRI learning mechanism,” Machine Learning, vol. 1, pp. 11-46,
Technical Memo, Menlo Park, CA, 1980. 1986.
I531 C. F. Cettys and T. A. Willke, “The application of Bayes’Theo- P. Nii, ”Blackboard systems: Blackboard application sys-
rem when the true data state i s uncertain,” Organizational tems,” A / Magazine, vol. 7, no. 3, Conference, pp. 82-106,1986.
Behavior a n d Human Performance, pp. 125-141, 1969. I. R. Benoit et al., ”ALLIES: An experiment in cooperating
[54] D. A. Schum, “Current developments in research on cas- expert systems for command and control,” in Proc. of the
caded inference,” in Cognitive Processes in Decision and Expert Systems in Gov’t Symp., pp. 372-380, 1986.
Choice Behavior, T. S. Wallsten, Ed. Hillsdale, NJ:Lawrence E. Durfee, V. Lesser, and D. Corkill, “Coherent cooperation
Erlbaum Press, 1980. among communicating problem solvers,”Techn. Rep. 85-15,
[55] C. W. Kelly, Ill,and S. Barclay, “A general Bayesian model for Department of Computer and Information Science, Univer-
hierarchical inference,” Organizational Behavior and Human sity of Massachusetts, Amherst, MA, pp. 2-8, 1985.
Performance, vol. I O , 1973. P. Nii, “Blackboard systems: The blackboard model of prob-
(561 J. Pearl, “On evidential reasoning in a hierarchy of lem solving,” A/ Magazine, vol. 7, no. 2, pp. 38-53, Summer
hypotheses,” Artificial Intelligence, vol. 28, no. 1, Feb. 1986. 1986.
[57l T. 0. Binford and T. S. Levitt, “Bayesian inference in model- D. R. Reddy, L. D. Errnan, and R. B. Neely, “The HEARSAY
based machine vision,” in Proc. AAA/ Uncertainty in Artificial speech understandingsystem:An exampleof the recognition
Intelligence Workshop, (Seattle, WA), July 1987. process,” in Proc. of the 3rd /]CAI, pp. 185-193, 1973.

1364 PROCEEDINGS OF THE IEEE, VOL. 76, NO. 10, OCTOBER 1988
 L. Erman, F. Hayes-Roth, V. Lesser, and D. Reddy, “The Hear- International Joint Conf. on Artificial Intelligence, (Karlsruhe,
say-ll speech-understanding system integrating knowledge West Germany), 1983.
to resolve uncertainty,” Computing Surveys, vol. 12, pp. 213- N. A. Khan and R. lain, ”Uncertainty management in a dis-
253,1980. tributed knowledge based system,” Proc. of the 9th Interna-
R. Balzer, L. Erman, P. London, and C. Williams, “HEARSAY- tional joint Conf. on Artificial Intelligence, (Los Angeles, CA),
I l l : A domain-independent framework for expert systems,” Aug. 1985.
in Proc. of the 1st National Conf. on AI, pp. 108-110,1980. S. Y. Lu and H. E. Stephanou, A Set-Theoretic Framework for
V. R. Lesser and D. D. Corkill, “The distributed vehicle mon- the Processingof Uncertain Knowledge, AAAI-84, Austin, TX,
itoring testbed: A tool for investigating distributed problem Aug. 1984.
solving networks,” A/ Magazine, vol. 4, no. 3, pp. 15-33, Fall W. Lukaszewicz, “General approach to nonmonotonic logic,“
1983. in Proc. of the 8th InternationalJoint Conf. on Artificial Intelli-
L. Gasser, C. Braganza, and N. Herman, “Implementing dis- gence, (Karlsruhe, West Germany), 1983.
tributed AI systems using MACE,“ in Proc. 3rd Conf. on AI T. Niblett, “Judgemental reasoning for expert systems,” in Proc.
Applications, pp. 315-320, 1987. o f the 9th International joint Conf. on Artificial Intelligence,
D. D. Corkill, “A framework for organizational self-design in (Los Angeles, CA), Aug. 1985.
distributed problem solving networks,” Tech. Rep. 82-33, H. Prade, “A synthgetic view of approximate reasoning tech-
Department of Computer and Information Science, Univer- niques,” in Proc. o f the 8th International Joint Conf. on Artifi-
sity of Massachusetts, Amherst, MA, pp. 17-20, 1982. cial Intelligence, (Karlsruhe, West Germany), 1983.
S. Cammarata, D. McArthur, and R. Steeb, “Strategies of J. R. Quinlan, “INFERNO: A cautious approach to uncertain
cooperation in distributed problem solving,” in Proc. of the inference,” Technical Note N-1898-RC, The Rand Corporation,
8th IJCAI, pp. 767-770, 1983. Santa Monica, CA, Sept. 1982.
R. G. Smith and R. Davis, ”Frameworks for cooperation in -, “Consistencyand plausible reasoning,” in Proc. ofthe8th
distributed problem solving,” IEEE Trans. Syst., Man, Cybern., International Joint Conf. on Artificial Intelligence, (Karlsruhe,
vol. SMC-11, no. 1, pp. 61-70, Jan. 1981. West Germany), 1983.
R. Davis and R. G. Smith, “Negotiation as a metaphor for dis- E. Rich, Default Reasoning as Likelihood Reasoning, AAAI-83,
tributed problem solving,” Artificial Intelligence, vol. 20, pp. Washington, DC, Aug. 1983.
63-109, 1983. C. Rollinger, “How to represent evidence-Aspects of uncer-
D. D. Corkill and V. R. Lesser, ”The use of meta-level control tain reasoning,” in Proc. o f the 8th InternationalJoint Conf. on
for coordination in a distributed problem solving network,” Artificial Intelligence, (Karlsruhe, West Germany), 1983.
in Proc. of the 8th IJCAI, pp. 748-756, 1983. G. Shafer, “Probability judgment in artificial intelligence and
D. D. Corkill, V. R. Lesser, and E. Hudlicka, “Unifying data- expert systems,” presented at the Conference on the Calculus
directed and goal directed control,” in Proc. of the National of Uncertainty in Artificial Intelligence, George Washington
Conf. on AI, pp. 143-147, 1982. University, Dec. 1984.
E. H. Durfee, V. R. Lesser, and D. D. Corkill, ”Increasing G. Shafer and A. Tversky, “Languages and designs for prob-
coherence in adistributed problem solving network,” in Proc. ability judgement,” Cognitive Science, July-Sept., 1985.
of the 9th IJCAI, pp. 1025-1030, 1985. L. Shastri and J. A. Feldman, “Evidential reasoning in semantic
P. Friedland, ”Knowledge servers-Applications of artificial networks: A formal theory,” in Proc. of the 9th International
intelligence to advanced space information systems,” pre- Joint Conf. on Artificial Intelligence, (Los Angeles, CA), Aug.
sentation at AlAA NASA Conf., Washington, DC, June1987. 1985.
D. B. Lenat, ”Overcoming the brittleness bottleneck,” Key- E. H. Shortliffe, Computer-BasedMedicalConsultation: MYCIN.
note presentation to the IEEE Computer Science3rd Conf. on New York, NY: American Elsevier, 1976.
Artificial Intelligence, Klssimmee, FL, Feb. 1987. T. M. Strat, Continuous Belief Functions for Evidential Reason-
P. Harmon, Ed., Expert System Strategies, vol. 2, no. 8, San ing, AAAI-84, Austin, TX, Aug. 1984.
Francisco, CA, August 1986. M. Sullivan and P. R. Cohen, “An endorsement-basedplan rec-
H. G. E. Kyburg, “Knowledge and certainty,” in Proceedings ognition program,” in Proc. of the 9th InternationalJoint Conf.
AAA/ Uncertainty Workshop, pp. 30-38,1986. on Artificial Intelligence, (Los Angeles, CA), Aug. 1985.
T. R. Thompson, “Parallel formulation of evidential-reasoning
theories,“ in Proc. o f the 9th lnternational Joint Conf. on Arti-
BIBLIOGRAPHY
ficial Intelligence, (Los Angeles, CA), Aug. 1985.
Further Reading in Uncertainty R. M. Tong, D. G. Shapiro, J. S. Dean, and B. P. McCune, “A
comparison of uncertaintycalculi in an expert system for infor-
B. G. Buchanan and E. H. Shortliffe, Rule-BasedExpertSystems. mation retrieval,” in Proc. of the 8th InternationalJoint Conf.
Reading, MA: Addison-Wesley, 1984. on Artificial Intelligence, (Karlsruhe, West Germany), 1983.
P. Cheeseman, ”A method of computing generalized Bayesian L. R. Wesley, “Reasoning about control: The investigation of
probability values for expert systems,” in Proceedings o f the an evidential approach,” in Proc. o f the 8th InternationalJoint
8th International Joint Conf. on Artificial Intelligence, Karls- Conf. on Artificial Intelligence, (Karlsruhe, West Germany),
ruhe, West Germany, 1983. 1983.
R. A. Dillard, “Computing probability masses in rule-based sys- A. P. White, “Predictor: An alternative approach to uncertain
tems,”Technical Document 545, Navel Ocean Systems Center, inference in expert systems,” in Proc. of the 9th International
San Diego, CA, Sept. 1982. Joint Conf. on Artificial Intelligence, (Los Angeles, CA), Aug.
D. Dubois and H. Prade, ”combination and propagation of 1985.
uncertainty with belief functions,” in Proceedings o f the 9th R. R. Yager, “Reasoning with uncertainty for expert systems,”
International Joint Conf. on Artificial Intelligence, (Los Ange- in Proc. of the 9th InternationalJoint Conf. on Artificial Intelli-
les, CA), Aug. 1985. gence, (Los Angeles, CA), Aug. 1985.
L. Friedman, ”Extended plausible inference,” in Proc. of the L. A. Zadeh, “A computational approach to fuzzy quantifiers
7th InternationalJoint Conf. on Artificial Intelligence, Vancou- in natural languages,” ComputersandMathematics, vol. 9,1983.
ver, British Columbia, Canada, 1981.
M. L. Ginsberg, “Does probability have a place in non-
monotonic reasoning!” in Proc. of the 9th International Joint Further Reading for Expert Systems
Conf. on Artificial IntelIigence, (Los Angeles, CA), Aug. 1985.
J. Gordon and E. H. Shortliffe, “A method for managing evi- - B. Buchanan, G. Sutherland, and E. Feigenbaum, ”Heuristic
dential reasoning in a hierarchical hypothesis space,” HPP 84- DENDRAL: a program for generating explanatory hypotheses
35, Stanford, CA, Sept. 1984. in organic chemistry,” in Machine lntelligence4, B. Meltzer and
J. Y. Halpern and D. A. McAllester, Likelihood, Probability, and D. Michie, Eds. Edinburgh University Press, 1969.
Knowledge, AAAI-84, Austin, TX, Aug. 1984. - R. 0. Duda and P. E. Hart, Pattern Classification and Science
J. H. Kim and J. Pearl, “A computational model for causal and Analysis. New York, NY: Wiley, 1973.
diagnostic reasoning in inference systems,” in Proc. of the8th - A. Hanson and E. Riseman, “VISIONS: A computer system for

FRANKLIN et al.: EXPERT SYSTEM TECHNOLOGY FOR THE MILITARY 1365

 interpreting scenes,” in Computer V/s/on Systems, A. Hanson Karl Keller received the B.S. degree in phys-
and E. Riseman, Eds. New York, NY: Academic Press, 1978. ics and engineering from Washington and
- B. P. McCune and R. Drazovich, “Radar with sight and knowl- Lee University, Lexington, VA, in 1983, and
edge,” Defense Electronics, vol. 15, no. 8, Aug. 1983. the M.E. degree in systems engineering
- H. Niemann and Y. T. Chien, Eds., “Knowledge based image from the University of Virginia, Charlottes-
analysis,” PatternRecognition (Special Issue)vol. 17, no. 1,1984. ville, VA, in 1985. He i s currently a Ph.D.
- L. Wesley and A. Hanson, “The useof an evidential-based model candidate in systems engineering at the
for representing knowledge and reasoning about images in the University of Virginia.
VISIONS system,” in Proc. of the Workshop on Computer He i s a member of the technical staff in
Vision: Representation and Control, Rindge, NH, Aug. 1982. the Artificial Intelligence Technical Center
oftheWashingtonC31DivisionoftheMlTRE
Further Reading for Maintenance and Troubleshooting Corporation, McLean, VA. He SI currently conducting applied
research in machine learning for military intelligence analysis. His
A. L. Brown, ”Qualitative knowledge, causal reasoning and the other research interests include knowledge representation,
localization of failures,” MIT AI Lab AI-TR-362, Ph.D. Thesis, knowledge acquisition, and knowledge-based simulation.
Nov. 1976.
R. Davis et al., “Diagnosis based on description of structure
and function,” in P ~ cAAA\-82,
. Aug. 1982.’ Tod S. Levitt (Member, IEEE) received the
J. DeKleer and J. S. Brown, ”Foundations of envisioning,” in B.S. degree from Case Western Reserve
Proc. AAAI-82, Aug. 1982. University, Cleveland, OH, and the M.A.
M. R. Genesereth, ”Diagnosis using hierarchical design and Ph.D. degrees in mathematics from the
models,” in Proc. AAAI-82, Aug. 1982. University of Minnesota.
J. J. King, “Artificial intelligence techniques for device trou- He i s currently a Principal Research Sci-
bleshooting,” Computer Science Laboratory Technical Note entist at Advanced Decision Systems,
Series CSL-82-9 (CRC-TR-82-004), Hewlett Packard, 1501 Page Mountain View, CA. He i s the principal
Mill Road, Palo Alto, CA 94304, Aug. 1982. investigator in the development of a knowl-
W. R. Simpson and H. S. Balaban, ”The ARINC research system edge-based image understanding surveil-
testability and maintenance program (STAMP),” in Proc. 7982 lance system for the interpretation of syn-
/E€€ Autotestcon Conf., (Dayton, OH), Oct. 1982. thetic aperture radar imagery under DARPA’s Advanced Digital
Radar Imagery Exploitation System (ADRIES)project. As part of this
research, he has co-developed a probabilistic certainty calculus
jude E. Franklin (Senior Member, IEEE)
that performs model-based Bayesian inference to accrue evidence
for the belief in system hypotheses. The system performs partial
received the B.S., M.S., and Ph.D. degrees matching, accounts for and represents multiple (conflicting) inter-
in electrical engineering from Catholic Uni-
versity of America, Washington, DC. pretations of the military situation, and also models the influence
of terrain and military tactics in the evidential accrual process.
He is the Senior Vice President/General
Other work in evidential reasoning includes the use of inductive
Manager of the Technology Division and
learning techniques for eliciting rules and probabilistic estimates
the Manager of Researchand Development
from domain experts, domain independent evidential accrual in
at PRCIGIS, McLean, VA. At PRC, he directs
model-based, open-ended systems, and machine learning for
R&D programs in Expert Systems, Natural
induction of rules and probabilities. He is also project supervisor
Language, Distributed Problem Solving,
Computer Security, Feature Extraction, for the Knowledge-Based Vision section of DARPA’s Autonomous
Software Engineering, and Information Systems Engineering. Prior Land Vehicle (ALV) program. The objectives of this research are to
to joining PRC, he was Manager of the Navy Center for Applied provide the basic visual modeling, prediction, and recognition
Research in Artificial Intelligence, located at the Naval Research capabilities for the perceptual systemof the ALV. This work involves
Laboratoryand aVice President of Applied Engineeringat MAR Inc. research in machine vision for understanding natural terrain. In
His experience includes the design, analysis, and management of the course of this work, he has developed a mathematical frame-
AI programs, computer systems, sonar systems, communication work for the representation of thevisual memoryof a mobile robot
systems, and acoustic signal processing systems. that allows navigation and guidance basedon visual events without
Dr. Franklin i s a Fellow of the Washington Academy of Sciences, the need of precise metric information. This work has been imple-
member of American Association of Artificial Intelligence, Sigma mented in the QUALNAV model for qualitative spatial reasoning.
X i Honorary Research Society, and Acoustics Society of America.
His current interests are in the fields of image understanding and
He i s the author of over 40 technical papers, conference papers, computer vision, artificial intelligence, evidential reasoning, and
and technical reports. He has been active in the organization of digital signal and image processing.
IEEE technical conferences including vice chairman for the 1987
and chairman of the 1989 Conference on AI Systems in Govern-
ment, co-editor of an IEEE EXPERTspecial issue on Applications of Brandon L. Buteau(Associate, IEEE)was born
AI for the Government. He serves on the AFCEA committee to in Boston, MA, on February 8, 1954. He
investigatetechnology insertion into SDI and a special committee received the B.S. degree cum laude in
on Security & Integrity of SDI. applied mathematics (computer science)
from Harvard University, Cambridge, MA,
in 1976.
Since then he has worked at Planning
Cora lackey Carrnody (Member, IEEE) Research Corporation (PRC), McLean, VA,
received the B.S. and M.A. degrees in math- on the design and development of infor-
ematicsfrom theJohnsHopkins University, mation systems and technology for a vari-
Baltimore, MD, and the M.S. degree in com- ety of national intelligence systems. His
puter science from Fairleigh Dickinson efforts have earned commendations from both the Defense Intel-
University, Rutherford, NJ. ligence Agency and the Defense Communications Agency. In his
She i s Chief Scientist of PRC/GIS’s Space current role as a systems applications scientist, he i s leading a
Systems Technologies Division, McLean, researchproject in the areaof distributed problem solvingfor PRC’s
VA, where she is currently supporting the Government Information Systems group. This project involves the
Space Station Software Support Environ- integration of several advanced technologies into a unified prob-
ment project for NASA‘s Johnson Space lem-solving architecture, including distributed expert systems,
Center. She has been with PRC since 1978, primarily in military natural language understanding, active temporal data bases, and
information systems, specifically command & control software. machine learning.

1366 PROCEEDINGS OF THE IEEE, VOL. 76, NO. IO, OCTOBER 1988

View publication stats