Engineering Applications of Artificial Intelligence 16 (2003) 91–103

Application of Bayesian decision networks to life cycle engineering in

Green design and manufacturing
J.Y. Zhu, A. Deshmukh*
Department of Mechanical and Industrial Engineering, University of Massachusetts, Amherst, MA, USA

Abstract

Environmental impact assessment of design and manufacturing decisions have received significant attention in the recent years.
Researchers have not only focused on industrial waste minimization and chemical substitution in processes or products, but also on
the effect of product design decisions on the environment during the manufacturing, in-use and end-of-life stages of the product.
This research investigates the applicability of Bayesian decision networks to study the impact of design decisions on the life cycle
performance, including environmental friendliness, of a product. Bayesian decision theory provides a normative framework for
representing and reasoning about decision problems under uncertainty. A framework for integrated analysis of the product life cycle
is presented. We discuss the specification of domain models for wide range of processes, such as manufacturing, recycling and
disposal, an action model, and an utility model.
r 2003 Elsevier Ltd. All rights reserved.

Keywords: Environmental impact analysis; Life cycle engineering; Bayesian decision theory; Decision networks

1. Introduction include raw material scrap, coolants, lubricants, solvents

and other catalysts used in the manufacturing processes,
Historically, companies have invested vast amounts of waste related to the energy used by the processes, and
resources in the development of manufacturing process the product disposal waste. Accurate estimation of the
technology, while paying relatively less attention to the environmental impact of a product or process is
environmental consequences of these developments. In important since it can influence public policy decisions,
recent years, the environmental issues have gained such as the polluter-pays principle or the allocation of
prominence as the severity of the impact of manufactur- environmental credits to various manufacturers.
ing on the global climate becomes evident (Cattanach In order to stay competitive, organizations need to
et al., 1995). satisfy the increasing customer demand for innovative
However, the task of estimating the environmental products. The growing complexity of products on one
impact of different manufacturing processes is extremely hand and shortening technology life cycle on the other
difficult. Fig. 1 shows the complexities involved in hand lead to increasing pressure on enterprises and
analyzing environmental impact throughout the life design engineers. Hence, minimal resources are devoted
cycle of a product. The raw material extracted from the to analyzing the tradeoffs between different waste
earth is transformed into a finished product via a series streams generated during the product life cycle, the
of processes in manufacturing plants. Aggregate waste impact of product design on these waste steams, and
streams are produced at each stage of the manufacturing linkages between environmental impact and maintaining
operations. After the product is sold to the customer, production rate to meet the market demand. Life cycle
additional waste streams may be created during the engineering, which presents a systematic approach to
period of service. At the end of a product’s useful life, it integrating design, manufacturing and operational
is re-manufactured, recycled or disposed off. The waste decisions related to a product with the objective of
streams at different stages of the product life cycle may optimizing the performance over the life cycle of a
product, has been proposed as an approach for reducing
*Corresponding author. Tel.: +1-413-545-1615; fax: +1-413-545- the overall harmful environmental impact of a product
1027. (Sincero and Sincero, 1996; Graedel and Allenby, 1995).

0952-1976/03/$ - see front matter r 2003 Elsevier Ltd. All rights reserved.
doi:10.1016/S0952-1976(03)00057-5
92 J.Y. Zhu, A. Deshmukh / Engineering Applications of Artificial Intelligence 16 (2003) 91–103

Oil waste
surfactant
flocculant CO2 H2S
antifoam
.......... Recycle
Gases
Wastes Purifica
tion

Flash
Drum
Raw Material Extraction

Raw Material Waste Energy Utilization Liquid and Solid Waste Recycling
Waste Minimization
Fig. 1. Major waste streams generated during a product life cycle.

In life cycle engineering, elements such as design, cycle time, poor quality and environmentally unfriendly
production, distribution, application, maintenance, dis- products.
posal, recycling and their environmental impact are The last few years have seen a surge in interest in
considered simultaneously. Thus, life cycle engineering Bayesian decision theory in the artificial intelligence field
presents a potential methodology for evaluating pro- using various graphical dependence models (Pearl,
ducts from an end-to-end perspective. It supports 1988). Bayesian decision theory provides a normative
processes to evaluate the resource consumption and framework for representing and reasoning with decision
environmental burden associated with a product or a problems under uncertainty (Jensen, 1996). In this
process. The evaluation is conducted by identifying and framework, a general probabilistic model of the problem
quantifying the energy and material usage and environ- domain is created. The model can be used to reason
mental releases across all stages of the product or the about the possible state of the problem domain given
process. partial observations of the domain. For each state of the
Interest in life cycle engineering has increased in domain, a set of alternative actions is also specified in an
recent years as design engineers have realized that they action model. For each state and each available action
need to integrate manufacturing cost considerations and in that state, the action model describes into which state
environmental concerns in the engineering design the problem domain may be transferred as a conse-
process (Thurston and Essington, 1993), and the use quence of the action and how probable is the
of even partial or uncertain information can enormously transformation.
reduce the product costs and environmentally hazardous To count the preference of the decision maker, a
waste. Hence, a decision support system capturing utility model is specified to describe the desirability of
imprecision and modeling uncertainties in early design each state of the domain. Once the domain model, the
and development stages of a product or process can be action model and utility model are specified, the optimal
an important tool in the life cycle analysis process. This decision or decision sequence under each state of the
decision support system should work concurrently with domain can be computed in the form of a policy.
knowledge-based engineering in a product design and Although a decision problem of a small size can be
manufacturing development framework. The tradeoffs solved using decision trees, it is computationally
between different alternatives, both in manufacturing expensive to solve a decision problem in a domain with
and downstream use, can be analyzed before focusing a large number of states, with a sequence of decisions to
on a specific process. This tool would give the design make, and with a large number of alternative actions per
and manufacturing engineers some degree of quantita- state. Recent advances in probabilistic reasoning explore
tive understanding about recycling, disposal and their structured representation of domain models. The con-
environmental impact. However, integrating decisions ditional independence among domain variables is
over the entire life cycle of a product presents a encoded in a network structure such that the joint
challenging task due to the different time scales and probability distribution of the domain can be specified
uncertainty associated with long term decisions. The by local distributions over small subsets of variables
current lack of technology for coordinating design (Pearl 1988; Xiang et al., 1993). This structure not only
decisions and managing change over the life of a simplifies the representation but also guides the in-
product often results in higher product costs, longer ference computation. Reasoning under uncertainty has
 J.Y. Zhu, A. Deshmukh / Engineering Applications of Artificial Intelligence 16 (2003) 91–103 93

been applied to various application domains, such as Denote the cardinality of N by n: Normally, accurate
preserving shoreline under climate uncertain climate characterization of a jpd requires the specification of
changes (Chao and Hobbs, 1997) and estimating effects parameters that are exponential order of n in number.
of petrochemicals (Cattanach et al., 1995). The major benefit of using a Bayesian network
Our investigation applies advances in Bayesian representation is that the jpd over a very large set of
decision networks to the engineering life cycle assess- variables can be compactly specified by a much smaller
ment problem. We investigate proper representations of number of variables. For example, if a domain can be
the problem domain that can provide sufficient aid in described by a set N of 20 binary variables, specifying a
decision making and are computationally tractable. general jpd requires about 220 ¼ 106 parameters. If the
First, we discuss the specification of the domain model, dependence relations of these variables can be repre-
the action model, and the utility model. This involves sented as a Bayesian network where each node has no
identification of design alternatives, state descriptors, more than 3 parents, then the jpd can be specified using
manufacturing alternatives, maintenance alternatives, only about 20 * 23 ¼ 160 parameters. The representation
recycling and disposal, etc. To be able to use existing task is simplified significantly. The resultant jpd,
computational tools which can handle discrete domains however, contains all the information that is needed to
effectively, we perform appropriate discretization for the answer any query of the form ‘‘what is the probability of
domain and decision variables. Once the models are X given the value of Y ?’’, where X and Y are any
specified, we investigate policy computation using subsets of N: The conditional independence in Bayesian
available computational tools. network structure plus the knowledge of the local
problem model allow us to specify the complete joint
distribution over the entire domain.
2. Overview of Bayesian decision networks In general, decision networks can be used to decide
the best action given partial observations of the world.
A Bayesian network consists of a directed acyclic A decision network represents the knowledge about an
graph (DAG) and an associated joint probability uncertain problem domain, the available actions at each
distribution (jpd). The nodes in the graph are labeled state of the domain and the desirability of each state.
by the random variables that define the jpd, we will Decision networks are extensions of Bayesian networks.
therefore use nodes and variables interchangeably in this A Bayesian network can be used to infer the state of the
paper. These random variables represent states of the world given partial observations of the world. In order
system. Each variable may take its value from two or to avoid exponential explosion in representation of the
more possible values. For example, a variable denoting uncertain domain, a structured representation, such as a
the intensity of a waste stream, may take its value Bayesian network, is often preferred. It consists of
from no-waste, low-waste, mild-waste, high-waste and chance nodes and arcs between them. The actions are
catastrophic. represented as decision nodes. The availability of an
The arcs in the DAG specify causal relations among action at different states of the domain is represented by
the variables. An arc from variable x to y indicates that having specific chance nodes as its parents. The
x is a direct cause of y: Using standard terminology in desirability of the states of the domain is represented
graph theory, x is the parent of y and y is the child of x: by value or utility nodes.
A directed path x-y-z indicates that y shields all the In order to explain the applicability of Bayesian
causal influence of x to z; in other words, z and x are networks, consider the following example: Suppose a
conditionally independent given y: If neither x nor y has company wishes to purchase a used pipe system to
any parent, the two variables are marginally independent. transport chemicals in its plant. The company finds a
That is, normally, they are not relevant to each other. system constructed of composite coiled-tube and priced
The dependence and independence relations repre- at$10 000. The company knows that if this type of
sented by the DAG allows the jpd to be specified locally coiled-tube system is in good shape, it is worth $11 000.
by the conditional probability distribution of each node It also learns that 80% of the coiled-tube systems were
conditioned on its parents (Zhu et al., 1998). Let N be manufactured in a plant which produces high quality
the set of all variables. Let pðxÞ be the parent nodes of x products, while 20% were manufactured in a plant
in the DAG. The conditional probability distribution of which produces inferior quality products, or lemons.
x conditioned on its parents is denoted by PðxjpðxÞÞ: Furthermore, it learns that of the ten coiled tube
When x has no parents, the corresponding distribution segments, each good system has one segment that will
is simply PðxÞ: The jpd PðNÞ over all variables can then fail, whereas each inferior system has six segments which
be written as will fail.
Y The pipe system has never been repaired and no
PðNÞ ¼ PðxjpðxÞÞ: segment of coiled tube has been replaced. Hence, it
xAN either has one segment out of ten that will fail or six
94 J.Y. Zhu, A. Deshmukh / Engineering Applications of Artificial Intelligence 16 (2003) 91–103

segments that will fail. A contractor has agreed to Note that a great deal of redundancy is present in the
charge $400 to repair one segment and $2000 to repair representation. In Fig. 2, a subtree starting with the
six segments. Suppose the company also learns that it decision of whether to purchase the pipe system is
has the option of performing some tests on the pipe repeated seven times. If we add another test option, it
system before purchasing it. Specifically, it can have one will be repeated at least two more times. Moreover, if we
segment tested, and the cost of the test is $90, or it can add a decision before the test options, the entire tree
have two segments tested at the cost of $130. Hence the may be repeated for each possible choice at that
company now has two decisions: first it must decide decision.
whether to perform tests; second, it must decide whether In this research, we demonstrate how to apply
to purchase the pipe system. Bayesian decision networks to large and complex
We can apply decision tree representation to solve this domain problems. A decision network consists of three
problem as shown in Fig. 2. Each branch from left to types of nodes:
right represents either an action (e.g., D0 ) or a possible
1. Zero or more chance nodes, which contain proposi-
consequence (e.g., G on upper right). Each consequence
tional variables. They are represented by circles in the
is labeled by its probability of occurrence (e.g., 0.8 for G
DAG.
on the top) and the utility (e.g., 600 for G). The value for
2. Zero or more decision nodes, which contain choices
consequence junction is calculated by weighted average
available to the decision maker. They are represented
(e.g., 600 * 0:8  1000 * 0:2 ¼ 480  200 ¼ 280). The va-
by squares in the DAG.
lue for action junction is the maximum value among
3. One value node, which contains a random variable,
that of the next level junctions (e.g., 280 ¼ maxð280; 0Þ).
whose value is the utility of the outcome. This node is
represented by a diamond in the DAG.
0.8 G
600
280 B 280 As in a standard causal network, the arcs into chance
D0 0.2 L
-1000 nodes show the variables upon which that node is
R conditionally dependent. The arc into decision nodes
0
show exactly which variable will be known to the
0.9 G
B 350 510 decision maker at the time the decision is made. The arcs
350 into the value node show which variables enter into the
T 0.1 L
-1090 calculation of the utility.
262 0.8 R
280 D1 -90 As shown in Fig. 3, the first decision node contains
0.4 G 510
the decision whether to perform the test, the possible
B -450
-90 values of decision are no test, test one, or both test one
0.2 0.6 L
-1090 and test two. The second decision node contains the
T R
-90 decision whether to buy or not to buy the pipe system.
0.96 G
Test one and test two are chance nodes. They contain
47 0
406 B 406 the results of the tests, such as test is not run, test is
0.04 L
-113 0 positive or the test is negative. The condition node is a
0.667 R chance node as well. It represents a good condition or a
T, S
-130
bad condition of the pipe system. The node Value V
0.6 G
-130 B -170 470 represents a function of the variables.
0.133 T, S 0.4 L For example, if decision node1 has value D1 ¼ t1 (the
-1130
R
first test), we will spend $90 to run test one. If decision
228 -130
D2
0.6 G
-130 B -170 470

0.133 T, S 0.4 L
-1130
R -130 Decision Decision Value
Test 1
0 G node node
47 0
-130 B -1130
1 L
-113 0
0.067 T, S R
-130
Test 2 Condi
tion
Fig. 2. The decision tree for the pipe system example. The symbols are
defined as follows: D0 : no test; D1 : test segment one; D2 : test segment
two; T: one segment test succeeds; :T: one segment test fails;
S: second segment test succeeds; :S: second segment test fails;
B: purchase; R: not purchase; G: high quality; L: lemon. Fig. 3. The decision network for pipe system example.
 J.Y. Zhu, A. Deshmukh / Engineering Applications of Artificial Intelligence 16 (2003) 91–103 95

node2 has the value D2 ¼ b1 then we buy the pipeline. If tion stored in the decision network. The required
the chance node C has value C ¼ c1 then the pipeline in conditional probabilities can be obtained in the follow-
good condition. The company makes $1000 when it ing way: Given D1 ¼ d1 and D2 ¼ b1 ; the three chance
purchases the pipeline and spends $400 to repair the nodes comprise a Bayesian causal network. Using the
pipeline. The expected value will be V ðd1 ; b1 ; c1 Þ ¼ inference computation for a Bayesian network, we can
90 þ 1000  400 ¼ 510: Actually, this node can be compute the probability of c1 given ðt11 ; t12 Þ and
considered a propositional variable which can take probability of c2 given ðt11 ; t12 Þ in that network.
exactly one value for each combination of values of its The main advantage of the decision network is to
parents. The possible values which a variable contri- reduce the redundancy in representation and to provide
butes to V can be stored on the arc for that variable. a compact and intuitive formalism for modeling the
Each chance node is conditionally dependent on its decision problem. Conditional independence embedded
parents as in a standard causal network. in the network structure specifies the full joint distribu-
In Bayesian networks, we also need to specify tion over domain by requiring the specification of only
prior probabilities of the root nodes and the condi- local distributions. Hence, any query that can be
tional probability of the non-root nodes. Suppose answered by a decision tree can be answered using a
PðgoodjpipelineÞ ¼ 0:8 and PðlemonjpipelineÞ ¼ 0:2 are decision network. Using shells for decision networks,
given. Then, Pðþt1 jgoodpipelineÞ ¼ 0:9 since a good the redundant components of the decision tree are
pipeline has one of 10 segments bad, and Pðþt1 j generated on the fly, and analysis and inference can be
lemonpipelineÞ ¼ 0:4 since a lemon pipeline has six out carried out automatically.
of 10 segments bad. Given that the pipeline is good and
the first test is positive, the probability that the second
test is positive is Pðþt2 jgoodpipeline; þt1 Þ ¼ 8=9 ¼ 0:889: 3. Bayesian networks for life cycle engineering
Given that the pipeline is bad and the first test is posi-
tive, the probability that the second test is positive is Traditionally, decisions involved in the life cycle of an
Pðþt2 jlemonpipeline; þt1 Þ ¼ 3=9 ¼ 0:333: All these va- engineering product were made independently. Design
lues are stored in the decision network as shown in Fig. 4. engineers did not anticipate potential production or
Now in order to make the decisions, we perform environmental problems. Their primary focus was on
inference in the decision network. For each decision engineering properties of the products. The economic
combination, we calculate its expected utility. First, we and environmental analyses were performed after the
order the decision nodes according to time to obtain the fact, after design and manufacturing alternatives had
ordering ½D1 ; D2 : Next we let D1 equal its first value d1 been developed. Life cycle engineering is a systematic
and proceed to next decision node D2 : When this approach integrating the design and manufacture of
decision is made, the information at T1 and T2 is products with the view of optimizing all elements
available. We can then compute the expected utility by involved in the life cycle of the product.
marking T1 and T2 together with a 1 and letting T1 Trade-offs between economic and technical consid-
equal its first value t11 and T2 equal its first value t21 : erations need be considered well before alternative
Then we let D2 equal its first value b1 ; and compute the configurations are developed, as it has been estimated
expected value of V given these instantiations. The that 70–80% of the cost of product development and
expected value EðV jd1 ; t11 ; t21 ; b1 Þ ¼ V ðd1 ; b1 ; c1 ÞPðc1 jd1 ; manufacture is determined at the initial design states.
t11 ; t21 ; b1 Þ þ V ðd1 ; b1 ; c2 ÞPðc2 jd1 ; t11 ; t12 ; b1 Þ; where the Life cycle engineering can improve efficiency signifi-
values of V can be obtained directly from the informa- cantly if combined with Bayesian decision theory which
provides a normative framework for representing and
P(+t1|good pipe-line) = 0.9 reasoning with decision problems under uncertainty.
Test 1 P(+t1|lemon pipe-line) = 0.4
This work presents a methodology for formulating
P(-t1|good pipe-line) = 0.1
P(-t1|lemon pipe-line) = 0.6 design optimization problems that directly incorporates
manufacturing, economic and environmental considera-
tions into the initial design decision-making process.
P(good pipe-line) = 0.8 Our representation using a decision network consists of
Test 2 Condi P(lemon pipe-line) = 0.2 four classes of variables:
tion
1. design, manufacturing, end of life use and other
P(+t2|good pipe-line, +t1) = 0.899 P(-t2|good pipe-line, +t1) = 0.111 choices,
P(+t2|good pipe-line, -t1) = 1 P(-t2|good pipe-line, -t1) = 0 2. performance measures,
P(+t2|lemon pipe-line, +t1) = 0.333 P(-t2|lemon pipe-line, -t1) = 0.677 3. desirability of products in terms of performance
P(+t2|lemon pipe-line, +t1) = 0.444 P(-t2|lemon pipe-line, -t1) = 0.566
measures, and
Fig. 4. The resultant decision network. 4. economic considerations in terms of costs.
96 J.Y. Zhu, A. Deshmukh / Engineering Applications of Artificial Intelligence 16 (2003) 91–103

The first class of variables are decision nodes. Design As the weather condition is not controllable, we
choices may include materials, dimensions, shapes, use expected load limit 80 * PðnormalweatherÞ þ
surface finish, etc. Manufacturing choices may include 50 * PðsevereweatherÞ: Using PðnormalweatherÞ ¼
process choices such as shaping, machining, assembling, 1  PðsevereweatherÞ; the expected load limit becomes
etc. End of life use choices may include re-manufactur- 80  30 * PðsevereweatherÞ: As PðsevereweatherÞ in-
ing, recycling and disposal. creases from 0 to 1, the expected load limit decreases
Performance measures of the product may include from 80 to 50. Formally, the expected utility of a
load limit, speed limit, vibration, life span, etc. The product based on a particular performance measure
actual performance of a product being designed usually vi is
cannot be precisely determined by the design/manufac- X
ui ¼ uðvi ¼ vij ÞPðvi ¼ vij jpi Þ;
turing choices. For example, how much traffic a bridge j
can endure before collapsing cannot be determined
precisely based only on its design. Considerable where pi is a set of choices and conditions that vi is
uncertainty exists due to many operational factors dependent upon (directly or indirectly), Pðvi jpi Þ is the
which are not under the control of the design engineer. performance distribution, and uðvi Þ is the utility
As a consequence of the uncertainty about the distribution. Similarly, the expected utility of a product
performance limit of products, over-design is common based on a particular cost hi is
X
in engineering products, as the design engineer tries to ci ¼ cðhi ¼ hij ÞPðhi ¼ hij jpi Þ;
place the normal work load of a product well below its j
actual performance limit. where pi is a set of choices and conditions that hi is
For example, suppose the normal load of a particular dependent upon, Pðhi jpi Þ is the cost distribution, and
product is distributed in the range of ð40; 80Þ measured cðhi Þ is the utility distribution of hi : The total expected
according to some unit. The expected value of the load is utility of a produce design is then
in the center of the range (around 60). An over design X X
Des1 may have the load limit in the range of ð90; 130Þ in u¼ s i * ui  s k * ck ;
i k
which the products under Des1 have their load limit
around the center of the range (about 110). Hence for where each si ðsk Þ is a coefficient which determines the
majority of products and majority of loads, there is relative weight of each performance measure.
about 50 units of unused load performance. The
consequence is that extra materials, energy, labor and
other resources are wasted to ensure the additional 50 4. Implementation
units of load performance. Due to the above reasons, we
represent performance measures as chance nodes, which We demonstrate the implementation of the concept
are dependent on choice nodes and possibly other using WEBWEAVR-III (Haddawy, 1999), a toolkit for
chance nodes. Bayesian decision networks.
Different performance measures and costs cannot be The major steps involved in the implementation are:
directly compared. For example, if a design Des2 has
1. Represent four classes of variables and their relations
20% higher load limit but 30% higher manufacturing
as a decision network using the Network Editor.
cost than another design Des1 : Which one should we
2. Compile the decision network into a runtime
choose? It certainly depends on the utility the user
structure using the Network Compiler.
places on higher load limit and the reduced cost. In
3. Inference about the performance distribution and
general, we evaluate the desirability of performance
cost distribution given a set of design/manufacture/
measures and costs in terms of utilities and represent
end of life use choices using the Inference Engine.
them as utility nodes. Utility nodes may have deci-
4. Calculate the total expected utility for each design
sion nodes and chance nodes as their parents. To
based on the above inference results.
simplify the representation and decision making, a
5. Repeat the last two steps for each design.
common assumption made is additivity: the utility of
6. Determine the optimal design for various design
individual product performance and cost can be
constraints.
combined linearly.
Due to the existence of chance nodes, we cannot
evaluate the desirability of a design deterministically, Our pilot implementation is performed on the design
but instead must compare different design options of an oil-drill platform. Drilling is a critical component
through their expected utility. For example, the load for oil production. The decision network for the
limit of a product is 80 units if the product operates platform design example is shown in Fig. 5. The
under normal weather conditions, but the load limit network defines all possible alternative processing paths
reduces to 50 units under severe weather conditions. in the life cycle for a given design. Four decision
 J.Y. Zhu, A. Deshmukh / Engineering Applications of Artificial Intelligence 16 (2003) 91–103 97

alternatives are represented by material choice, manu- the network shown in Fig. 6 is identified by a name, such
facturing choice, endoflifeuse choice and quantity as material or cost1 or limit-load, and a unique integer,
choice. Each choice is specified by a design variable. specifying the node number. Each node is assigned the
Material choice has the alternatives of steel, alumi- node number based on the sequence in which it was
num and composite. Manufacturing choice has the created in the network. Hence, material, 6 denotes
alternatives of casting, milling and shaping. Endoflifeuse material node which is identified internally by WEB-
choice has the options of re-manufacture, recycling and WEAVER as node 6 and this node has three alter-
disposal. Quantity choice has the options of large, natives, steel, aluminum and composite.
medium and small, each of which can be associated with For the performance measures, the designers can
a numeric range. Each choice or decision has an choose the criteria of load limit, speed limit and waste
associated cost distribution. Note that each node of stream. Each performance measure is specified by the
three values of high, medium and low as shown in Fig. 6.
material Every performance measure has an associated utility
manufacture distribution over its possible values conditioned on its
parents, the choice variables. In practice, these distribu-
tions can be obtained by estimation from available
endoflifeuse statistics or from subjective estimation of domain
limit_load limit_speed
experts. The probability distribution associated with
wastestream speed limit is shown in Fig. 7. Estimation of these
probability distributions is critical in any application as
it can drastically affect the predictions and outcomes of
utility1 utility2 cost2 the analysis.
The probability distributions for performance mea-
utility3 sures, selecting different design choices for the example
cost 1 are shown in the following figures. In Fig. 8, the designer
cost3
chooses steel for material, milling for manufacturing, re-
quantity
manufacturing for endoflifeuse, and a large production
Fig. 5. The decision network for life cycle analysis of an oil-drill. quantity.

Fig. 6. The prior probabilities for oil-drill example. Each variable is labeled by its name followed by a variable index, e.g., material, 6 and utility2, 1.
98 J.Y. Zhu, A. Deshmukh / Engineering Applications of Artificial Intelligence 16 (2003) 91–103

Fig. 7. The probabilities associated with speed limit.

Fig. 8. The event probabilities for choosing re-manufacturing as endoflifeuse.

Marginal probability distributions for all variables are particular example is 3.1433. The total expected utility
then computed for this design. We can then calculate the of each possible design can be computed by a simple
expected utility of the design. The value for this program outside of WEBWEAVR-III. The optimal
 J.Y. Zhu, A. Deshmukh / Engineering Applications of Artificial Intelligence 16 (2003) 91–103 99

Fig. 9. The event probabilities for choosing recycling for endoflifeuse.

configuration can be identified based on the expected all variables are then automatically computed. The
total utility from the various partial design alternatives. designer thus directly accesses performance, quality,
In Fig. 9, the designer changes the endoflifeuse waste, cost and utility of a product up front. This
variable value from re-manufacturing to recycling. It is facilitates reduction of the lead time of product,
assumed (reflected in the utility distributions) that increases product development efficiency and reduces
recycling induces more cost on the product itself than the operation wastes.
re-manufacturing, but it will reduce aggregated waste at Table 1 shows the experimental results for 18 different
each stage. A new set of the probability distributions are design choices, their performance measures and ex-
calculated. The expected utility is computed as 3.5533. pected utilities obtained by our implementation. The
Note that the expected utility increased from 3.1433 to first three columns show the design choices for material,
3.5533. It indicates that the second decision is better end of life use, and quantity. The next nine columns are
than the first one. divided into three groups. Each group represents one of
In Fig. 10, the designer modifies the endoflifeuse the performance measures in load limit, speed limit and
variables value from recycling to disposal. It is assumed waste stream. Each group has three columns represent-
that disposal induces the highest expenditure on waste ing the probability distribution over the possible values
management. We obtain a new set of the probability of the performance. The last column shows the expected
distributions and a new value of utility 3.3933. The value utility of the design. This method enables us to analyze
is lower than that for recycling. Hence, in this example the often complex relationships among materials,
recycling results in the highest overall utility. One can manufacturing processes, waste management, as well
deduce that in this example the waste reduction at each as the economics of manufacturing operation.
stage outweights the cost savings associated with As an example of the computational process, consider
disposal. the second row of Table 1, where the design choices are
The designer can choose different combination of steel for material, recycling for end of life use, large for
material, manufacturing process, endoflifeuse. Each quantity, and mill for manufacture. After these choices
choice also includes the waste aggregated during the are entered into the WEBWEAVR-III inference engine,
production process and the probability distributions on it produces the distribution on performance measures as
100 J.Y. Zhu, A. Deshmukh / Engineering Applications of Artificial Intelligence 16 (2003) 91–103

Fig. 10. The event probabilities for choosing disposal for endoflifeuse.

Table 1
Experimental results

Design choice Performance measure ExpUtil

Material EndLife Quant LimitLd LimitSd WasteStr Value

High Med Low High Med Low High Med Low

Steel Remanu Large 0.67 0.25 0.08 0.8 0.1 0.1 0.3 0.5 0.2 3.41

Steel Recycl Large 0.67 0.25 0.08 0.8 0.1 0.1 0.1 0.3 0.6 3.55
Steel Dispos Large 0.67 0.25 0.08 0.8 0.1 0.1 0.1 0.2 0.7 3.39
Alum Remanu Large 0.6 0.2 0.2 0.08 0.6 0.32 0.3 0.5 0.2 2.71
Alum Recycl Large 0.6 0.2 0.2 0.08 0.6 0.32 0.1 0.3 0.6 2.85
Alum Dispos Large 0.6 0.2 0.2 0.08 0.6 0.32 0.1 0.2 0.7 2.69
Compos Remanu Large 0.7 0.15 0.15 0.2 0.4 0.4 0.3 0.5 0.2 1.32
Compos Recycl Large 0.7 0.15 0.15 0.2 0.4 0.4 0.1 0.3 0.6 1.46
Compos Dispos Large 0.7 0.15 0.15 0.2 0.4 0.4 0.1 0.2 0.7 1.3
Steel Remanu Small 0.67 0.25 0.08 0.8 0.1 0.1 0.3 0.5 0.2 0.26
Steel Recycl Small 0.67 0.25 0.08 0.8 0.1 0.1 0.1 0.3 0.6 0.3
Steel Dispos Small 0.67 0.25 0.08 0.8 0.1 0.1 0.1 0.2 0.7 0.14
Alum Remanu Small 0.6 0.2 0.2 0.08 0.6 0.32 0.3 0.5 0.2 0.81
Alum Recycl Small 0.6 0.2 0.2 0.08 0.6 0.32 0.1 0.3 0.6 0.85
Alum Dispos Small 0.6 0.2 0.2 0.08 0.6 0.32 0.1 0.2 0.7 0.69
Compos Remanu Small 0.7 0.15 0.15 0.2 0.4 0.4 0.3 0.5 0.2 1.72
Compos Recycl Small 0.7 0.15 0.15 0.2 0.4 0.4 0.1 0.3 0.6 1.76
Compos Dispos Small 0.7 0.15 0.15 0.2 0.4 0.4 0.1 0.2 0.7 1.60
 J.Y. Zhu, A. Deshmukh / Engineering Applications of Artificial Intelligence 16 (2003) 91–103 101

shown in the columns 4 through 12. For instance, the 5. Compilation and inference computation
distribution for load limit is (high: 0.6667, media: 0.25,
low: 0.0833). As specified in the network parameters, the In the previous section, we have demonstrated the use
conditional distribution of utility1 is as follows: of Bayesian decision analysis for design of an oil-drill
platform. We have not, however, explained how the
toolkit performs the necessary computation. In this
Util1 LimitLd pðUtil1jLimitLdÞ
section, we outline the key steps which ensure the
1 High 1.0 inference to be performed both correctly and effectively
0.8 Media 1.0 when the decision problems are large in size.
0.3 Low 1.0 In general, the topology of a decision network
contains (undirected) loops. For example, the structure
in Fig. 11 contains three loops. Effective exact inference
Based on this information and the distribution for cannot be performed in such structures (Xiang and
load limit, WEBWEAVR-III inference engine will Lesser, 2000). Instead, we need to compile the network
compute the posterior distribution of utility1 as into a tree to support such inference.
ð0:3 : 0:08333333; 0:8 : 0:25; 1 : 0:6666666Þ: The first step of compilation is to convert the DAG
into an undirected structure by pairwise connecting
The expected utility for the performance of load limit is
parents of each node and drop the direction of links.
then calculated as a weighted sum,
The resultant structure is called a moral graph of the
0:3 * 0:08333333 þ 0:8 * 0:25 þ 1 * 0:6666666 ¼ 0:89: DAG. The addition of links between parents of each
Similarly, the posterior distribution of utility2 is node is necessary since when a child variable is observed,
its ‘causes’ (its parents) will compete to explain the
ð0:4 : 0:1; 0:7 : 0:10000002; 1 : 0:8Þ observation, hence the dependence between the parents.
and that of utility3 is These links signify the dependence. Fig. 11 shows the
moral graph of the previous decision network.
ð0:2 : 0:10000001; 0:6 : 0:29999998; 1 : 0:6Þ: It has been shown (Jensen, 1996) that effective
They produce expected utilities of 0.91 and 0.8. inference can be performed by message passing in a
Similar to the conditional distribution of utility1 tree structure called a junction tree (JT) as illustrated in
above, the network model also specifies the condi- Fig. 12. Each cluster in the JT is labeled (inside) by a set
tional distribution for each of the cost variables. The of variables indicated by the index of each variable. A
inference engine will accordingly compute the posterior JT has the following running intersection property: the
distribution of each of cost1; cost2 and cost3: Through intersection of any two clusters is contained in each
a similar computation process, their expected utilities cluster on the unique path between them. For example,
are calculated as 0.1 for cost1; 0.3 for cost2; and 0.2 cluster C4 and C5 has the intersection C4 -C5 ¼ f9g:
for cost3: Using the weights for utility1; utility2 and Hence variable 9 is contained in the two clusters C3 and
utility3 as 2.0, 2.0 and 1.0, respectively, and the weights C8 on the path between C4 and C5 :
for of cost1; cost2 and cost3 as 2.0, 1.5 and 1.0, To compile an undirected graph into a JT, we must
respectively, the total expected utility of the design define the clusters. Each cluster in the JT should be a
alternative is maximal set of nodes that are pairwise connected in the
undirected graph, called a clique. Such a JT exists if and
2:0 * 0:89 þ 2:0 * 0:91 þ 1:0 * 0:8  2:0 * 0:1  1:5
only if the undirected graph is chordal. A path or loop in
* 0:3  1:0 * 0:2 ¼ 3:55:

Fig. 11. The DAG and corresponding moral graph of the decision network.
102 J.Y. Zhu, A. Deshmukh / Engineering Applications of Artificial Intelligence 16 (2003) 91–103

example, a message sent from C4 to C8 is a distribution

over variables 7 and 9.
Each inference consists of two passes of message
exchange. A cluster is selected as the root of message
passing, say, C5 : In the first pass, messages flow towards
the root as shown in Fig. 13 by the black arrows. After
the first pass is completed, a second pass starts where
messages flow away from the root as shown by the white
arrows. After the second pass, correct posterior prob-
abilities for each variable can be obtained from any
cluster containing it. The inference computation is linear
on the number of clusters and is exponential on the
cardinality of the largest cluster. Hence it can be
efficiently used for large scale environmental decision
Fig. 12. The junction tree. problems if the largest cluster is small in size.

c5
5, 8, 9 6. Summary
c3
c9
Life cycle engineering at the design stage can
3, 6, 9 8, 12
significantly reduce cost and environmental impact over
c2 the life of a product. Automated decision aids will
c6 c8 facilitate rational decision making in concurrent en-
2, 12
6,10 6, 7, 9 gineering design. However, direct representation of all
design alternatives and their evaluation is computation-
c7
c0 ally intractable. We explore the structured represen-
c4 6, 7, 11
0, 10 tation which allows effective representation and
4, 7, 9
decision-making. A methodology using Bayesian deci-
c1 sion networks for life cycle engineering is proposed. A
1, 11
pilot implementation on an oil-drill design using
WEBWEAVR-III was performed. We can observe that
Fig. 13. Inference by message passing in the JT. in this example the recycling alternative results in
highest overall utility, although the disposal option
would result in lower short term costs. The ability of the
an undirected graph has a chord if there is a link in the proposed approach to incorporate long term implica-
graph between two non-adjacent nodes on the path or tions in presence of uncertainty in design decisions is
loop. A undirected graph is chordal if every loop of important in estimating true costs of engineering design
length longer than 3 has a chord. Therefore, in order to decisions.
compile the moral graph of a decision network into a The proposed method is founded in Bayesian
JT, the moral graph must be made chordal by adding probability theory and decision theory, and a rigorous
some links, called fill-ins. The moral graph in Fig. 11 is theory on graphical representation of probabilis-
already chordal, but in general, a moral graph may not tic dependency. The computations in this analysis
be chordal. are exact. Therefore, unlike ad-hoc methods, the
Once the JT is constructed, the probability distribu- proposed method introduces no systematic decision
tions of the decision network can be converted to errors. The reliability of the results depends mainly on
distributions associated with each cluster. For each two factors: the accuracy of the graphical dependence
variable in the decision network, its associated distribu- structure and that of the associated distributions. It has
tion is assigned to a cluster in the JT if the cluster been shown that the accuracy of the structure is more
contains the variable as well as its parents in the decision crucial than that of the distributions (Jensen, 1996;
network. Haddawy, 1999).
Effective inference can be performed by passing Future research directions include application and
messages along the links of the JT. Each link in testing of our method in complex product design
the JT is associated with the intersection of the two situations and development of a system to facilitate
clusters it connects. A message over a link is a the preference and utility specification process based on
distribution over the corresponding intersection. For the inputs from the domain experts.
 J.Y. Zhu, A. Deshmukh / Engineering Applications of Artificial Intelligence 16 (2003) 91–103 103

References Sincero, A.P., Sincero, G.A., 1996. Environmental Engineering: A

Design Approach. Prentice-Hall, Englewood Cliffs, NJ.
Cattanach, R.E., Holdreith, J.M., Reinke, D.P., Sibik, L.K., 1995. Thurston, D.L., Essington, S.K., 1993. A tool for optimal
Handbook of Environmentally Conscious Manufacturing. Irwin manufacturing design decisions. Manufacturing Review 6 (1),
Press, Homewood, IL. 48–59.
Chao, P.T., Hobbs, B.F., 1997. Decision analysis of shoreline Xiang, Y., Lesser, V., 2000. Justifying multiply sectioned Bayesian
protection under climate change uncertainties. Water Resources networks. Proceedings of the 4th International Conference
Research 33 (4), 817–830. on Multi-agent Systems, Boston, MA, January, 2000. pp.
Graedel, T.E., Allenby, B.R., 1995. Industrial Ecology. Prentice-Hall, 349–356.
Englewood Cliffs, NJ. Xiang, Y., Poole, D., Beddoes, M.P., 1993. Multiply sectioned
Haddawy, P., 1999. An overview of some recent developments in Bayesian networks and junction forests for large knowledge based
Bayesian problem-solving techniques. AI Magazine 20 (2), 1–4. systems. Computational Intelligence 9 (4), 171–220.
Jensen, F.V., 1996. An Introduction to Bayesian Networks. UCL Zhu, J.Y., Cooke, W.B., Xiang, Y., 1998. Application of Bayesian
Press, London, UK. networks to quantified risk assessment. Proceedings of Fifth
Pearl, J., 1988. Probabilistic Reasoning in Intelligent Systems: International Conference on Industrial Engineering and Manage-
Networks of Plausible Inference. Morgan Kaufmann, Los Altos, ment Science, Beijing, pp. 321–328.
CA.