Agents: definition and formal architectures

CS7032: AI for IET

Lecturer: S Luz
[email protected]
September 23, 2014

Outline

What is an agent?

Properties one would like to model

Mathematical notation
A generic model

A purely reactive model

A state-transition model
Modelling concrete architectures
References: (Wooldridge, 2009, ch 1 and ch 2 up to section2.6), (Russell
and Norvig, 2003, ch 2)

Abstract architectures will (hopefully) help us describe a taxonomy (as it

were) of agent-based systems, and the way designers and modellers of multi-
agent systems see these artificial creatures. Although one might feel tempted to
see them as design tools, abstract architectures are really descriptive devices.
Arguably their main purpose is to provide a framework in which a broad variety
of systems which one way or another have been identified as agent systems
can be classified in terms of certain formal properties. In order to be able to
describe such a variety of systems, the abstract architecture has to be as general
as possible (i.e. it should not commit the modeller to a particular way of, say,
describing how an agent chooses an action from its action repertoire). One
could argue that such level of generality limits the usefulness of an abstract
architecture to the creation of taxonomies. This might explain why this way of
looking at multi-agent systems hasnt been widely adopted by agent developers.
In any case, it has the merit of attempting to provide precise (if contentious)
definitions of the main concepts involved thereby moving the discussion into a
more formal domain.

1
What is an agent?

A person who acts on behalf of another (Oxford American Dictionaries)

Action, action delegation: state agents, travel agents, Agent

Smith
Action, simply (autonomy)

Artificial agents: understanding machines by ascribing hu-

man qualities to them (McCarthy, 1979) (humans do that all
the time)
Interface agents: that irritating paper-clip

An example: Interface agents

Embodied life-like characters REA: real estate agent (Cassell, 2000), speech recognition and

synthesis, multi-modal interaction (gestures, gaze, life-size graphics etc)

Not all UI designers like the idea, though (Shneiderman and Maes, 1997)

AIs Nobel prize winner

Scientific and philosophical background:

Natural Sciences
Conformance to natural laws
An air of necessity about them
Sciences of the artificial (Simon, 1981):
Prescription vs. description
A science of the contingent
A science of design

2
From a prescriptive (design) perspective...

Agent properties:
Autonomy: acting independently (of user/human intervention)
Situatedness: sensing and modifying the environment
Flexibility:
re-activity to changes
pro-activity in bringing about environmental conditions under
which the agents goals can be achieved, and
sociability: communicating with other agents, collaborating, and
sometimes competing

In contrast with the notion of abstract architecture we develop below, these

properties of multi-agent systems are useful from the point of view of conceptu-
alising individual systems (or prompting the formation of a similar conceptual
model by the user, in the case of interface agents). Abstract architectures, as
descriptive devices, will be mostly interested in the property of situatedness.
The starting point is the assumption that an agent is essentially defined by
its actions, and that an agents actions might be reactions to and bring about
changes in the environment. Obviously, there is more to agents and environ-
ments than this. Agents typically exhibit goal-oriented behaviour, assess future
actions in terms of the measured performance of past actions, etc. The ar-
chitectures described below will have little to say about these more complex
notions.
However, before presenting the abstract architecture, lets take a look at a
taxonomy that does try to place goals in context:

PAGE Descriptions
P.A.G.E.: Percepts, Actions, Goals and Environment

Agent Type Percepts Actions Goals Environment

Symptoms, find-
Medical diagnosis Questions, tests, Healthy patient,
ings, patients Patient, hospital
system treatments minimize costs
answers

Satellite image Pixels of varying Print a categoriza- Correct categoriza- Images from orbit-
analysis system intensity, color tion of scene tion ing satellite

Pixels of varying Pick up parts and Place parts in cor- Conveyor belt with
Part-picking robot
intensity sort into bins rect bins parts

Print exercises,
Interactive English Maximize students
Typed words suggestions, cor- Set of students
tutor score on test
rections

As in (Russell and Norvig, 1995)

...alternatively, P.E.A.S. descriptions

3
 Performance mea-
Agent Type Environment Actuators Sensors
sure
Symptoms, find-
Medical diagno- Healthy patient, Questions, tests,
Patient, hospital ings, patients
sis system minimize costs treatments
answers

Satellite image Correct catego- Print a catego- Pixels of varying

Satellite link
analysis system rization rization of scene intensity, color

Pick up parts
Part-picking % parts in cor- Conveyor belt Pixels of varying
and sort into
robot rect bins with parts intensity
bins

Print exercises,
Interactive Students score Set of students;
suggestions, cor- Keyboard input
English tutor on test testing agency
rections

From (Russell and Norvig, 2003)

Environment properties
Fully vs. partially observable: whether agents can obtain complete and accurate in-
formation about the environment
deterministic vs. stochastic: whether the next state of the environment is fully deter-
mined by the current state and action performed by the agent
episodic vs. sequential: whether agents next action depends only on the current state
of the environment (episodic), or on assessment of past environment states (sequential)
static vs. dynamic: whether the environment changes independently of the agents
actions
discrete vs. continuous: whether the possible actions and percepts on an environment
are finite (discrete environment) or not (continuous environment)
single vs. multiple agents

Types of environments
Environment Observable Deterministic Episodic Static Discrete Agents
Crossword puzzle fully yes sequential static yes single
Chess w/ clock fully strategic sequential semi yes multi
Poker partially strategic sequential static discrete multi
Backgammon fully stochastic sequential static discrete multi
Car driving partially stochastic sequential dynamic continuous multi From (Russell and
Medical diagnosis partially stochastic sequential dynamic continuous single
Image analysis fully deterministic episodic semi continuous single
Robot arm partially stochastic episodic dynamic continuous single
English tutor partially stochastic sequential dynamic discrete multi
Plant controller partially stochastic sequential dynamic continuous single
Norvig, 2003)

Chess with clock is semi-static (semi-dynamic) because even though the

environment changes (time passes, players are penalised if they fail to play by a
certain deadline) independently of the agents actions, it does so in a regular and
predictable way. The environment a taxi driver operates in, on the other hand,
may change in unpredictable ways, being therefore a dynamic environment.
Medical diagnosis typically depends on patient history, so its environment is
sequential. In contrast, a robot which examines various parts which pass on a
conveyor belt in front of the robots camera, one at the time, and moves defective
parts off the conveyor belt does not have to consider its past decisions in order
to decide whether the part it is currently examining is defective. The robot is
therefore said to operate in an episodic environment.

Abstract agent architectures

Why?
What we would like to describe:

4
 agent
environment

Agent

Perception Action

Environment
their interactions

Notation

A few tools from discrete maths and logic:

A, S, ... : sets
(S) : the powerset of S

S : all sequences of elements (i.e. ordered subsets) of S
, , , : material implication, conjunction, disjunction and
negation
Quantifiers: and

Refreshing your memory: The powerset of a set S is a set containing all of

Ss subsets. So, if S = {s0 , s1 }, its powerset (S) = {, {s0 }, {s1 }, {s0 , s1 }}.
Sequence sets S are similar to power sets except that the order of their elements
matters. So, if S = {s0 , s1 }, we have S = {<>, < s0 >, < s1 >, < s0 , s1 >, <
s1 , s0 >}.
The logical connectives behave in the familiar way. X Y is true iff both
X and Y are true; X Y is true if X is true, Y is true, or both; X Y is
false if X is true and Y is false, and true otherwise. Since these connectives are
used here simply to illustrate particular examples of agent architectures, we will
deliberately leave their semantics underspecified. Feel free to interpret them
as they are interpreted in, say, PROLOG. One could also add existential and
universal quantifiers as needed.

Defining an architecture
A standard architecture is a 4-tuple: Archs =< S, A, action, env >
where
S = {s1 , s2 , ..., sn } is a finite set describing all (possible) environment
states, and
A = {a1 , a2 , ..., an } is the set of all actions an agent is capable of
performing
action is a function describing the agents behaviour, and
env is a function describing the behaviour of the environment

An example: consider the blocks world below:

5
 c
a b
000000000000000000000000000000000000
111111111111111111111111111111111111
000000000000000000000000000000000000
111111111111111111111111111111111111

Assume the robotic arm is capable of performing a set of actions A = {move(a, b), move(c, f loor), . . . }.
Environment states could be represented by an S = {s1 , s2 , ..., sn }, where we
could have
s1 = {lef t(a, b), on(f loor, a), on(a, c), . . . } (1)
and (after a move(c, f loor)):
s2 = {lef t(a, b), on(f loor, a), on(f loor, c), . . . } (2)
One could also, for instance, encode general properties such as
x, y, z.lef t(x, y) lef t(y, z) lef t(x, z) (3)
etc, to remain constant in all environment states.

Agenthood
An agents behaviour will be characterised by the following function:
action : S A
Does having S as the domain of action make the function most naturally
suited to modelling episodic or sequential environments? Why?
Requirement captured: the current action may depend on the interaction
history (i.e. the sequence of environment states)
Actions consider an essentially sequential environment, since their arguments
are sequences. In our blocks world example, one could have action(hs1 , s2 i) =
move(a, b), assuming that the purpose of moving block c off block a was to be
able to place a on b. Note, however, that the formalism does not explicitly
represent purpose, planning, or any such notion.

Environment dynamics
Changes in the environments will be characterised by the following function:

env : S A (S)

Intuition: env(sj , ak ) = S 0 performing an action ak on an environment whose

state is sj results in a number of scenarios (S 0 )
In the general case, what type of environment does env model? Answer: env
models a non-deterministic environment.
If |S 0 | = 1, then the environment is deterministic.
In the example above, if nothing else is said, action move(c, f loor) would
affect the environment as follows:
env(s1 , move(c, f loor)) = {{lef t(c, a), . . . }, {lef t(a, c), . . . }} (4)
If the relative positions of the blocks were irrelevant, env(s1 , move(c, f loor))
would be a singleton.

6
Interaction history
The agent-environment interaction will be characterized as follows:
a a au1 a
u
h : s0 0 s1 1 ... su ...

h is a possible history of the agent in the environment iff :

u N, au = action(< s0 , ..., su >) (5)
and
u > 0 N, su env(su1 , au1 ) (6)
Condition (5) guarantees that all actions in a history apply to environment
states in that history.
Condition (6) guarantees that all environment states in a history (except the
initial state) result from actions performed on environment states.

Characteristic behaviour
Characteristic behaviour is defined as a set of interaction histories

hist = {h0 , h1 , ..., hn }

where
each hi is a possible history for the agent in the environment

Invariant properties
We say that is an invariant property of an agent architecture iff

For all histories h hist and states s S

is a property of s (written s |= )

We will leave the relation |= underspecified for the time being.

In architectures where agents have minimal symbolic reasoning abili-
ties (e.g. some concrete reactive architectures), |= could be translated
simply as set membership, that is s |= s, in architectures
where reasoning isnt performed

Behavioural equivalence
Equivalence of behaviours is defined as follows (in an abstract architecture)

An agent ag1 is regarded as equivalent to

agent ag2 with respect to environment envi
iff hist(ag1, envi ) = hist(ag2, envi )

When the condition above holds for all environments envi , then we simply
say that ag1 and ag2 have equivalent behaviour

7
Modelling reactive agents

Reactive architectures
production rules
a scheme for defining priorities in the application of rules (e.g. sub-
sumption (Brooks, 1986))
P ...
e A
r c
Wander
c
e t
p Explore i
t o
i Move forward
n
o
n Avoid Obstacles

no reasoning (theorem proving or planning) involved

Abstract reactive agents

Purely reactive agents can be modelled by assuming

action : S A

Everything else remains as in the general (standard abstract architecture)

case:

Archr =< S, A, actionr , env >

Abstract reactive agents operate essentially on an episodic view of envi-

ronments (i.e. they are memoryless agents). See slide 6

Example: a thermostat.
temperature(cold) do(heater(on))
temperature(cold) do(heater(of f ))
S = {{temperature(cold)}, {temperature(cold)}}
A = {heater(on), heater(of f )}

Purely reactive vs. standard agents

Proposition 1 (exercise):

Purely reactive agents form a proper subclass of standard

agents. That is, for any given environment description S,
and action repertoire A:
(i) every purely reactive agent is behaviourally equiva-
lent to a standard agent, and
(ii) the reverse does not hold.

8
Modelling Perception
Refining the agents decision function

see action

Agent

Perception Action
Environment

Types and sources of perception

Percepts and actions

Facts perceived by an agent will be represented as set

P = {p0 , p1 , ..., pn }

The decision function becomes

action : P A

which is then linked to environment states via the perception fumction

see : S P

Properties of perception
If see(si ) = see(sj ), we say that si and sj are indistinguishable, even if
si 6= sj .
Define , an equivalence relation over S by saying s s0 iff see(s) =
see(s0 )
If | | = |S|, then we say that the agent is perceptually ominiscient
On the other hand, if | | = 1, then the agents perceptual ability is nil
In the following example, four environment states are (rightfully, assuming
that a butterfly doesnt cause changes in the weather) as two:
x = weather(cold)
y = moving wings(butterf ly)
S = {{x, y}, {x, y}, {x, y}, {x, y}}
see({x, y}) = see({x, y})
see({x, y}) = see({x, y})
The equivalence relation is then {{{x, y}, {x, y}, }{{x, y}, {x, y}}}
For perceptually omniscient agents, see(si ) = si , for all si S. In an agent
with no perceptual abaility, see(si ) = see(sj ) for all si , sj S.

9
Refining the interaction history

Representation used so far: history as a sequence of environment states

see Agent
action

next state

Perception Action

Environment

The next step: represent history (in the agent architecture) as environment
changes as perceived by the agent

State-based decision function

The state-based architecture will be represented as

Arch =< S, A, P, I, action, env, see, next >

where
I is the set of all internal states of an agent,
see : S P ,
action : I A, and
next : I P I

Properties of State-based architectures

Proposition 2:

State-based architectures are equivalent to standard

architectures with respect to the behaviours they
are able to represent.

Exercise

Prove Proposition 2.

Further information

(Weiss, 1999): overview; this presentation mostly followed the material

found there;

(Russell and Norvig, 2003): agent = architecture + program

(Genesereth and Nilsson, 1987): foundations of a theory of rational agency

10
References
Brooks, R. A. (1986). A robust layered control system for a mobile robot. IEEE
Journal of Robotics and Automation, RA-2(1):1423.

Cassell, J. (2000). Embodied conversational interface agents. Communications

of the ACM, 43(4):7078.
Genesereth, M. and Nilsson, N. (1987). Logical Foundations of Artificial Intel-
ligence. Morgan Kaufmann, Los Altos, CA.
McCarthy, J. (1979). Ascribing mental qualities to machines. Technical Memo
326, Stanford University AI Lab.
Russell, S. J. and Norvig, P. (2003). Artificial Intelligence. A Modern Approach.
Prentice-Hall, Englewood Cliffs, 2nd edition.
Shneiderman, B. and Maes, P. (1997). Direct manipulation vs interface agents.
interactions, 4(6):4261.
Simon, H. A. (1981). The Sciences of the Artificial. MIT Press, Cambridge,
Massachusetts, second edition.
Weiss, G. (1999). Multiagent Systems. MIT Press, Cambridge.

Wooldridge, M. (2009). An Introduction to MultiAgent Systems. John Wiley &

Sons, 2nd edition.

11