Department of Computer Science© G.M.P O'Hare University College Dublin DEPARTMENT OF COMPUTER...

Department of Computer Science© G.M.P O'Hare

University College Dublin

DEPARTMENT OF COMPUTER SCIENCE

Robotics Part I & IIMulti-Agent Systems(MAS)

G.M.P. O'Hare


The Turing TestAllan Turing [5] in his classic paper ‘Computing Machinery and Intelligence’, circumvented the problem of defining artificial intelligence. Such a test took for form of a game.

The game he describes has three participants, an interrogator, a human and a machine. The interrogator is physically removed from the other two participants. He can communicate with each of them by way of a teletype, he does not however, know which participant is machine and which is human.

His task is to establish which one is the machine and which is the human. This became renowned as the ‘Turing Test’.

A computer could be thought to display intelligence if the interrogator could not distinguish between man and computer.


The Turing Test IITuring’s work did not, however, win universal acceptance. More recently opponents like Millar [6] while recognising the merits of his work highlights the fact that it does not yield any insight into the various skills which constitute intelligence.

He believed this to be of great significance if any realistic attempt is to be made at constructing a truly intelligent machine.

If I may paraphrase Leonardo de Vinci (1452-1519), he in a similar vein suggested that.........

“when man understands thenatural flight of the bird, man will be able to build a flying machine.”


A Working Definition

So with artificial intelligence, the definition we shall employ is that volunteered by Marvin Minsky [7]

“Artificial intelligence is the science of making machines do things that would require intelligence if done by man.”


A Simple Example

E 4 F 7

"If there is a vowel on one side of a card thenthere will be an even number on the other"

How many cards must you turn over in orderto test the validity of this statement?


Another Simple Example

1

2 3

4 5 9

7

8 6

Players alternatively choose a card until they select three cards in total

The Object of the Game is to obtain a total of 15 and ensuring your opponent does not acquire a total of 15.

What strategy would you adopt?


The History of AI 1

The term Artificial Intelligence is normally attributed to John McCarthy.

In 1956 he organised a conference which was to enable researchers in the field to share expertise. As a consequence of his actions the discipline of AI was founded.

Some attendees namely, Allan Newell, Herbert Simon and Marvin Minsky himself, are now without question the leading researchers in the field.


The History of AI 2

At the conference Newell & Simon detailed work on the theorem prover Logic Theorist which had been performed at Carnegie.

This is commonly regarded as the first AI program as such. The Logic Theorist was written in IPL (Information Processing Language) the first language which permitted computers to process concepts as opposed to numerical quantities.


The History of AI 3

Minsky & McCarthy founded the MIT AI Laboratory.

McCarthy is renowned as the inventor of LISP while Minsky proposed the Frame concept for Knowledge Representation.

In this early stage efforts tended to concentrate on:

Game PlayingGame Playing: equipping a computer to play a particular game.

Theorem ProvingTheorem Proving: equipping a computer to show that some statement follows logically from a set of known truths called axioms.


The History of AI 4

Early efforts employed a technique known as State Space SearchState Space Search involving essentially several components ...

(a) an initial stage

(b) a final state

(c) an ability to detect final state

(d) a set of legal operations that can be applied to each state.

Such an approach can often be understood better by conceptually regarding states as nodes and operations as arcs.


The History of AI 5

By way of example in a chess game:

(a) initial state: initial state of chess board.

(b) final state: checkmate.

(c) ability to detect final state: ability to detect checkmate.

(d) set of legal operations: legal moves of chess.


Generate & Test 1

The simplest form of state space search is that of Generate & TestGenerate & Test.

Such an approach involves typically three stages, those of ...

(a) Generating a possible solution in the form of a new state.

(b) Ascertaining whether the new state is indeed the final state.

(c) If new state is the final state terminate, otherwise repeat steps a, b and c.


Generate & Test 2Two forms of generate and test exist: Depth-first SearchDepth-first Search & Breadth-first SearchBreadth-first Search.

Both fall foul of the ‘combinatorial explosion’, caused by the exponential growth of the nodes irrespective of the order of generation.

Consequently exhaustive search is only feasible when the search space is very small.

For larger spaces the search needs to be guided.

Guided searches are normally referred to as Heuristic SearchesHeuristic Searches. Searches of this nature utilise domain specific knowledge called heuristics.


Guided Searches

Guided searches are normally referred to as: Heuristic SearchesHeuristic Searches. Searches of this nature utilise:

domain specific knowledge called heuristics.


Exercise 1

Attempt to draw a state space for the famous missionaries and cannibals problem


Distributed Artificial Intelligence

Distributed Artificial Intelligence(DAI) :-Endeavours to achieveIntelligent Systems not by constructing a large Knowledge-Based System, but rather by partitioning the knowledge domain and developing 'Intelligent Agents',each exhibiting expertise in a particular domain fragment.

This group of agents will thereafter collectively work towards the solution of global problems.


The Co-operating Experts MetaphorThis solution of problems by a group of agents, providing mutual assistance as and when necessary is often referred to as the.....

"Community of Co-operating Experts Metaphor"

Smith and Davis, Lenat, Hewitt

Proponents of this philosophy believe that reciprocal co-operation is the cornerstone of society.


Social Agents

3

4

6

2

5

Q?

R AND P -> Q

R 4P 6

R?

P?M 2.4

L 5S 4

M -> PL OR S -> M

S?L?

M?

AquuaintanceModel

Domain SpecificKnowledge Base


Why Distributed Artificial Intelligence? Mirrors Human Cognition

Potential Performance Enhancements

Elegantly Reflects Society

Incremental Development

Increased Robustness

Reflects Trends in Computer Science in General

Strong Analogies to Decompositional Techniques employed in Software Engineering


Coordination Paradigms

Numerous Different Paradigms have been proposed....

The Blackboard Model (Reddy et al)

The Actor Model (Hewitt)

The Contract Net Approach (Smith and Davis)

The 'BEING' Concept (Lenat)

Hybrid Approaches


The Blackboard Model

DAI first presented itself in the form of a blackboard model in the HEARSAY I,II and III (Carnegie Mellon Univ) speech understanding systems

The Blackboard model involves agents communicating by way of a shared global data structure called the 'Blackboard'

Agents could not communicate directly with each other but only via the contents of the blackboard.


Poll-Hypothesis-TestThe Model generally involves a poll-hypothesis-test cycle. In the case of the HEARSAY systems this consisted of....

Initially a hypothesis is installed on the Blackboard regarding the meaning of a particular utterance

poll :- each ks is polled in order to ascertain if it can refine the. hypothesis Those which can indicate their ability to do so

with an associated confidence factor

hypothesis :- The ks boasting the highest cf is invoked making the appropriate refinement to the hypothesis on the blackboard.

test :- The other kss evaluate the amendment and based on the collected response the amendment is either adopted or

ignored

The hypothesis is successively refined repeating this cycle until the utterance is identified with a sufficient degree of confidence


Blackboard ProblemsDAI approaches seem to have followed a course similar to those developments which have taken place in the design of real-time languages

The blackboard exhibits obvious similarities to Hoare's Monitor concept and in general the Mail-box concept

It also suffers from the same limitations :-

• Congestion Problems

• Reliability Problems

Thus there was a realisation as in real-time language design that....

"To divorce data transmission and process synchronisation was totally unnatural"

Young, Real-Time Languages


The Actor ModelHewitt's Actor Model was one such example of later synchronised approaches

The society of co-operating experts were to be modelled by a basic building block called an Actor

Actors could communicate with other Actors via messages

The behaviour of an actor was to be defined dependent upon which message it received and these potential actions are contained in a Script

Actors could communicate with those actors with whom they were acquainted as indicated by way of their acquaintance list

This model therefore offered point to point communication


The Contract Net Protocol I

The Contract Net Approach regarded the distribution of tasks amongst

agents, or the connection problem as a process of negotiation.

Contract Net Protocol:-

A node(the manager) advertises a problem via a broadcast to all other

nodes(potential contractors). Potential contractors compare this and other

problems and upon identification of the tasks for which they are most

suited they submit bids. The manager evaluates bids and awards contract

accordingly


Contract Net II

The communication protocol is vastly superior in that both manager and potential contractor participate in the decision regarding a suitable contract

It is also worth noting that here three addressing modes are offered:general broadcast, limited broadcast and point to point.

This approach adopts an inter agent language which while simple is capable of supporting the relevant communication.


BeingsLenat commissioned a very different approach when trying to overcome problems of inter agent understanding.

In his PUP6 system an agent was represented by a 'Being' which had to comply with a predefined structure.

It consisted of a fixed number of 'parts', each part representing a question that the ks may be equipped to answer.

If a part contained a value then the being is sufficiently knowledgeable with the value representing a procedural attachment which would yield the appropriate answer.

When a being asks a question it must therefore stipulate the relevant part.

This approach made question answering relatively trivial - essentially pattern matching. It avoided the need for an inter node language, however in so doing it enforced a very stylised form on each agent

Furthermore there was no provision for gaining expertise.


Hybrid Approaches

Hybrid Approaches

- Lesser and Corkill 1981

- Huhns et al 1983


Benevolence & Competition

Within all of these approaches there is this underlying presumption that the intelligent agents necessarily want to co-operate.

"The Benevolent Agent Assumption"

More recently a school of thought believes that this is not necessarily the case and agents may have conflicting goals

This has resulted in ....

"The Conflicting Agents Assumption"

Geneserth and Rosenchien Stanford HPP Project


Problems with DAI

• Identification of appropriate task decomposition and task distribution strategies

• Optimisation of problem solution (Cammarata et al 1982,1983)

• Difference of opinion between experts where the mapping between expertise and experts is not 1: 1 but 1: n - need conflict resolution strategies

• Problems with understanding

• Handling uncertainty becomes even more problematic

• Need deadlock avoidance strategies

• Problems with heterogenous nodes

• Interoperability


Reactive v Classical Systems

Essentially Multi-Agent systems occupy a point on a continuum between two extreme classes of system. These two extremes are...

• The classical system

• The reactive or situated action system

We propose a compromise that of the

'Deliberate Social Agent'


Percieve the world Represent the world

Reactive

Contemplative

Rea

son

ing

X

Complexity

X

X

Classical System

Deliberate Social System

Reactive

Situated Action System

Internal Model


Reactive or Situated SystemsAgents react to varying situations and consequently do not have an explicit representation of the world within which they exist.

Reasoning takes place within each agent at a very low level, essentially each agent has little more than an ability to perform pattern matching.

A given situation is characterised and matched against a collection of rules specifying appropriate behaviour associated with each of these situations ie situation -action or situated action.

Typically the actions associated with a given situation are often very simple and consequently the agents themselves are very simple computational entities.

Even though each of the individual agents are very simple the global complexity and global structures can be achieved as a result of the emergent property of the interacting behaviours of the community of agents.


Reactive Systems AssessmentAdvantages

* simplicity. * avoidance of necessity for a sophiciated representation of the world and more significantly the problems of maintaining this model. * generally the structure of agent interaction is well defined and domain independent.

Disadvantages

* New sets of rules need to be designed for each application. * Each situation needs to be specified and identified so as to have an associated rule.

* Difficulty in solving inherently recursive problems. * Lack of a precise theory upon which the combining behaviours of agents can be based and explained.


Reflective Systems

Generally the agents within a reflective system are more complex computational entities.

They do not merely react to a given situation in a specific way. In fact they may often react in different ways dependent on their own ‘beliefs’ or ‘intentions’.

Such systems necessitate an internal representation of the world. They often base their reasoning on the actions of the other agents within the community.

They normally possess some model of intentionality which represents their goals, desires, prejudices, beliefs etc. about themselves and the remainder of the community.

Certain classes of problem seem to necessitate this ability to reason using intentionality. The ‘wisest man’ puzzle seems to typify these.


Reflective Systems II

Reasoning intentionally normally demands use of higher order logics.

Particularly Modal logics.

- epistemic logics - doxastic logics

There are two general approaches

Sentential logics (Konolidge) Possible World Logics (Kripke)


Deliberative Systems Assessment

Advantages A clear theoretical reasoning model that underpins the approach;A mental state that is verifiable and traceable;Amenable to modeling using higher order logics;

DisadvantagesTheoretical Model is complex and unwieldy;Approach is more computationally demanding;Less appropriate for time critical reasoning scenarios;Necessitates the maintenance of a model of the environment;


The Cognitive ChasmExotic, FormalComplex Computationally Intractable Logicsof Intention

Pragmatic, Shallow,Simpistlic Multi-agentImplementation testbedswith little conformance to complex theoretical models

The 'cognitive chasm'that this thesis is seeking to bridge


Agents & The Notion of Agency

The term agent is somewhat nebulous and means different things to disparate research communities within Computer Science.

I use this terminology in the manner associated with the Distributed Artificial Intelligence (DAI) community, namely that agents are characterised by the attributes of autonomy; social ability; reactivity and pro-activity.

In addition a stronger notion of agency is often applied which demands that agents are ascribed mentalistic attitudes typically knowledge; belief; intention and obligation.

Wooldridge & Jennings (1995) distinguish between two usages of the term 'agent': the first is 'weak'; the second is stronger and potentially more contentious.


The Weak Notion of AgencyThe Weak notion of agency refers to the general way in which

the term agent is used to denote software (usually) or harware-based computer systems that has the following properties:

autonomy: agents operate without direct intervention with control over their actions and internal state;

social ability: agents interact with other agents and possibly humans via an agent-communication language;

reactivity: agents perceive their environment and respond to changes that occur within it in a timely fashion;

pro-activeness: agents do not simply respond to their environment, they are able to exhibit goal-directed behaviour (initiative).


The Stronger Notion of Agency

Within AI research, agency generally implies that in addition to the properties already outlined an agent is either conceptualised or implemented using concepts more usually applied to humans.

These properties include:

mentalistic notions of belief, knowledge, intention and obligation.


Other Attributes of Agency

Other attributes discussed in the context of agency include:

mobility: the ability to move around an electronic network;

veracity: the assumption that an agent will not knowinglycommunicate false information;

benevolence: is the assumption that agents do not have conflicting goals and that every agent will try and do what is asked of it.

rationality: in a crude sense the assumption an agent will act in order to achieve its goals, and will not act in such a way as to prevent its goals being achieved.


Definition of an Agent

Agents are often physically and logically distinct and are typically capable of reasoning, planning, communicating and cooperating (Hern 1988).

They may be robotic, be defined in terms of sensory input, motor control and time pressures, they may perform cognitive functions, react to stimuli, contain symbolic plans, or possess natural language capabilities.

Shoham (1993):“An agent is an entity whose state is viewed as consisting of mental components such as beliefs, capabilities, choices, and commitments.”


Belief Desire IntentionArchitectures

One particular Architecture that has been employed in the development of Reflective Systems is that of the Belief Desire Intention (BDI) Architecture.

The term BDI is attributed to Rao and Georgeff (1992).

The architecture models the reflective process in terms of theinterplay between these three mental attitudes.


Beliefs, Desires and Intentions

Let us consider these three mental attitudes:-

Belief : represents the information state of the agent, those things it believes to be true at a given instance;

Desire: represents the evaluative state of the agent that is those things that the agent at a given instance desires to bring about;

Intentions: represents those activities which the agent has decided at some previous time are crucial in achieving its goals in an adequate or optimum manner;


Early DAI Environments ABE (Erman et al 1988)

ARCHON (Wittig 1989)

CooperA (Sommaruga et al 1989)

MACE (Gasser et al 1987)

MADE (Wooldridge & O'Hare 1990)

Agent Factory (O'Hare 1992)

MCS (Doran et al 1991)

GBB (Hayes-Roth et al 1988)


Classes of Commitment

Elsewhere in the literature [RG92] it is recognised that varying degrees of commitment may be exhibited by agents.

Rao and Georgeff [RG91], [RG92] identify three discrete points on this commitment continuum, namely:

Blind Commitment,

Single-Minded Commitment and

Open-Minded Commitment.


Blind Commitment

Blind commitment is defined as the adherence to a commitment until such time as the agent believes it has achieved the commitment.


Single-Minded Commitment

Single-minded commitment represents a relaxation of blind commitment in that the agent will not drop its commitments unless it believes that they are no longer achievable.

The computational overhead of assertaining whether a given goal is achievable can be considerable.

Rao and Georgeff suggest that this can be achieved by permitting belief revision but not goal revision.


Open-Minded Commitment

Open-minded commitment offers a further relaxation in that an agent is willing to revise its goals and beliefs, retaining commitments that are still compatable with its goals.


Communication within DAI

Communication is central to the development of any satisfactory Multi-Agent System. Effective communicationis a prerequisite for achieving system coordination and system coherence.

Werner [Wer89] has identified several discrete classes of communication that occurs within Multi-Agent Systems.

These are:-

1. Complete abscence of communication;2. Inter-Agent Signalling;3. Message Passing;4. Plan Passing;5. Speech Acts;


Coordination

Coordination represents the problem or activity of reconciling the actions of the individual agent with those of the group or indeed organisation.

Since agents actions are derived from their goals and sinceagents are frequently benevolant their will inevitably beconflict.

Such interference can only be reconciled through communication.


Coherence

System coherence involves ensuring that the overall system performance is satisfactory


Absence of Communication

Sometimes communities of agents can achieve coherent behaviour without explicit communication.

Geneserth Ginsberg & Rosenchein [GGR84] considered this very issue in a seminal paper entitled Cooperation without Communication.

Agents might have a prearranged regime for achieving their goals and this is established a priori thus avoiding any need for dynamic communication.

Alternatively they may infer each others plans based on observations to date. This results in a prediction of agents'behaviour [Ros85].


Agent Signalling

Inter-Agent activity can be sychronised through the use of semaphore based technologies.

Semaphores offer a relatively simplistic communication technique.

They utilise the standard, primitives of wait and signal andare directly analogous to those techniques used within thedesign of real-time languages and systems


Message Passing

Another very common means of inter-agent communication is that of message passing.

Early work by Hewitt & Agha formulated a computationalparadigm based upon actor-based computation. Central to this was the notion of message passing.

Message passing generally manifests itsself in many DAI systems.


Plan PassingThis approach involves agents exchanging plans to one another. By so doing agents can anticipate the future directed actions of other agents.

One particular approach involves the exchange of Partial Plans. This approach called Partial Global Planning (PGP)was expounded by Durfee and Lesser. Within PGP agents build partial and incomplete plans which they subsequently share to colleagues in order to identify potential improvements.

This approach unlike for example multi-agent planning allowsagents to interleave planning and actions. Thus based uponfuture plans received agents can revise their plans and subsequently perform actions based upon this. PGPwas employed with great effect in the DVMT system.


Essence of Speech ActsThe origins of Speech Act Theory can be traced to thework of Austin [Aus62].

Two central characteristics associated with the basic theory of Speech Acts are:-

1. That human utterances are viewed as actions in a manner similar to physical operations that result in the movement of a book for example. They too result in a change in the state of the world.

2. That communication can be homogenised into a finite set of Speech Verbs that can be used to as an effective medium within which to communicate.


Speech Acts and State Change

It is not immediately obvious how Speech Acts result in a change to the environment.

All utterances are viewed as being situated within a particularcontext and each results in a revision to that very context.

The context is often viewed as the aggredation of the mental states of the participants namely the speaker and the hearer.

Such a mental state includes their Beliefs, Desires and Intentions.


A Pragmmatic Theory of Speech

We can thus view a pragmatic theory of speech as a function which takes a set of all utterances of a given language lets say L and an associated set of Contexts within which these can be expressed lets say C and derives the new context.

Thus

Speech_Function : L x C -> C


Speech Act ActionsAustin also identified three discrete classes of action associated with any given utterance:-

• Locutionary Acts :- which is performed by simply uttering a syntactically correct phrase;• Illocutionary Acts :- which is performed via a performative verb examples include tell, inform, ask, instruct, demand. Each verb has an associated illocutionary force. Austin identified some 1,000 such verbs in English;• Perlocutionary Acts:- is the bringing about of an effect on the hearer of the utterance;

Speech acts generally refer to the illocutionary act.


Speech Acts and Austin

Austin noted that certain utterances involved not merely the assertain of facts but rather the performance of associated action(s). These utterances are termed performatives and he noted that these like physical actions are prone to failure.

The conditions that must exist for sucessful completion werecalled felicity conditions. Three key conditions are:-

1. There must be an accepted procedure for the performative and the circumstances and individuals must be specified for this procedure.2. This procedure must be executed correctly and completely.3. The act must be performed in a sincere manner and any associated or implied behaviour honoured.


A DAI Textbook

Foundations of Distributed Artificial Intelligence, O'Hare, G.M.P., and Jennings, N.R., (Eds.), Wiley Interscience, 1996, ISBN 0-471-00675-0.

597 pages

Available Now

Department of Computer Science© G.M.P O'Hare University College Dublin DEPARTMENT OF COMPUTER...

Documents

Transcript of Department of Computer Science© G.M.P O'Hare University College Dublin DEPARTMENT OF COMPUTER...