MultiAgent Systems

Dr Oscar Lin

Traditional AI• Much traditional AI has been concerned

with how an agent can be constructed to function intelligently, with a single locus of internal reasoning and control implemented in a Von Neumann architecture.

• But intelligent systems do not function in isolation --- they are at the very least a part of the environment in which they operate, and the environment typically contains other such intelligent systems.

• Thus, it makes sense to view such systems in societal terms.

Outlines

• What is MAS?• What are the benefits of using MAS?• What are the challenges of developing MAS?• What are suitable domains for using MAS?

What is MAS?

• An extension of the agent technology where a group of loosely connected autonomous agents act in an environment to achieve a common goal.

• This is done either by cooperation where knowledge is shared among agents, or competition where knowledge is not shared.

• Demo: http://www.youtube.com/watch?v=ejhg0FpBaq0&feature=related

What is MAS?• Multiagent Systems

combine autonomous entities, each having diverging interests or different information.

• This comprehensive overview of the field offers a computer science perspective but also draws on ideas from … Computer

Science

Game theoryEconomics

Operations research

Logic

PhilosophyLinguistics

Benefits of Using MAS (motivations) • Distributed computations are sometimes easier to understand and easier to develop,

esp. when the problem being solved is itself distributed. • Distribution can lead to computational algorithms that might not have been

discovered with a centralized approach.• There are also times when a centralized approach is impossible, because the systems

and data belong to independent organizations that want to keep information private and secure for competitive reasons.

• The rationale for interconnecting computational agents and expert systems is to – enable them to cooperate in solving problems, – to share expertise, – to work in parallel on common problems, – to be developed and implemented modularly, – to be fault tolerant through redundancy, – to represent multiple viewpoints and the knowledge of multiple experts, and – to be reusable

Challenges (1)

• Environment– In a MAS, the action of an agent not only modifies its own

environment but also that of its neighbors.– This necessitates that each agent must be able to predict the actions

of other agents in order to decide the optimal action that would guide it towards the optimal goal.

– This type of concurrent learning could result in non-stable behavior and can possibly cause chaos.

– The problem is further complicated if the environment is dynamic.– Each agent needs to differentiate between the effects caused by the

actions of other agents and variations in the environment itself.

Challenges (2) • Perception – In a distributed multi-agent systems, the agents

are scatted all over the environment. – Each agent has a limited sensing capability

because of the range and coverage of the sensors connected to it.

– Therefore, decisions based on partial observations made by each of the agents could be sub-optimal and achieving a global solution by these means become intractable.

Challenge (3)• Abstraction

– In an agent system, it is assumed that an agent knows its entire action space, and mapping of the state space to action space can be done by experience. However, in MAS, every agent does not experience all of the states.

– To create a map, it must be able to learn from the experience of other agents with similar capabilities or decision making powers. In the case of cooperating agents which have similar goals, this can be done easily by establishing communication between the agents.

– In the case of competing agents, it is not possible to share information as each agent tries to increase its chances of winning.

– It is therefore essential to quantify how much local information and how much of the capabilities of the other agents must be known in order to create an improved model of the environment.

Challenges (4) • Conflict resolution– Conflicts stem from the lack of global view

available to each agent– Actions selected by an agent to modify a specific

internal state may be bad for another agent.– Under these circumstances, information on the

constraints, action preferences and goal priorities of each agent must be shared with other agents to improve cooperation. However, a major problem with MAS is determining when and to which agent to communicate this information to.

Challenges – (5) Inference

• Single agent system– Inference can be easily drawn by mapping the State Space

to the Action Space based on trial and error methods.

• MAS– It is difficult as the environment is being modified by

multiple agents that may or may not be interacting with each other.

– Further, the MAS might consist of heterogeneous agents, which are agents with different goals and capabilities.

Characteristics of Multiagent Environments

• Provides an infrastructure specifying communication and interaction protocols

• Are typically open and have no centralized designer

• Contain agents that are autonomous and distributed and may be self-interested or cooperative.

Agent Communications

• Coordination• Dimensions of meaning• Message types• Communication levels• Speech acts• KQML• KIF• Ontologies • Other communication protocols

Agent Communications

• Coordination• Dimensions of meaning• Message types• Communication levels• Speech acts• KQML• KIF• Ontologies • Other communication protocols

Agent Communications

• Why communicate?– In order to achieve better the goals of

themselves or of the society/system in which they exist.– Enable the agents to coordinate their

actions and behavior, resulting in systems that are more coherent.

Agent Communications

• Coordination–Coordination is a property of a system of

agents performing some activity in a shared environment.– The degree of coordination is the extent to

which they avoid extraneous activity by • reducing resource contention, • avoiding livelock and deadlock, and •maintaining applicable safety conditions.

Agent Communications• Coordination– Different ways of coordination• Cooperation: non-antagonistic

agents• Negotiation: competitive or

simply self-interested agents– To coordinate successfully, each

agent must maintain a model of the other agents, and also develop a model of future interactions. ----- sociability!

Coordination

Competition

Negotiation

Cooperation

Planning

Centralized planning

Distributed planning

Agent Communications• Coordination

– How a multiagent can maintain global coherence without explicit global control? – Coherence: how well a system behaves as a unit.– How?

• Some form of organization• Social commitment • Economic principles of markets

– Coherence and Optimality • are intimately related • Market mechanism – less effective in computing optimal allocations of resources

(Simon, 1996). • Organizational structures are essential for computing optimal allocations of

resources. – Herbert Simon. The Sciences of the Artificial. MIT Press, Cambridge, MA, third

edition, 1996.

Agent Communications --- Speech Acts• What is Speech acts?

– Speech act theory views human natural language as actions, such as requests, suggestions, commitments, and replies.

• Why do we need Speech Acts?– Goal: to model communication among computational agents using

human communication.– To insure that there is no doubt about the type of message.

• Problem: – in communication among humans, the intent of the message is not

always easily identified.

I am cold

I am cold

An assertion

A request for a sweater

A demand for an increase in room temperature

Agent Communications --- Speech Acts• A speech act has three aspects:

– Locution ---- the physical utterance by the speaker– Illocution ---- the intended meaning of the utterance by the

speaker– Perlocution --- the action that results from the locution

Physical sound/or text message

If all goes well, the window being shut

Intent for the message as q request or a

demand

John, please close the windows.

Agent Communications --- Speech Acts• Speech act theory uses the term performative to

identify the illocutionary force (言外之意 ) of this special class of utterance:– Example performative verbs: Promise, report, convince,

insist, tell, request, and demand

• Illocutionary force can be broadly classified as – assertives (statement of fact), – directive (commands in a master-slave structure), – commissives (commitments), – declaratives (statements of fact), and – expressives (expressions of emotions)

Agent CommunicationsKnowledge Query and Manipulation Language (KQML)

• KQML is a protocol for exchanging information and knowledge.• Structure:

(KQML-performative:sender <word>

:receiver <word> : reply-with : in-reply-to

:language <word> e.g. KIF, Prolog, LISP, SQL :ontology <word> define the common concepts, attributes, and relationships for different subsets of world knowledge :content <expression>

… )• Syntax

– Lisp-like. – Arguments identified by keywords preceded by a colon --- may be given in any order– The KQML-performatives: speech act performatives.

• Semantics – KQML-performatives: Domain independent – Message: defined by the fields

Parameters of the message passing

The semantics of the message

Agent Interaction Protocols

• Interaction protocols govern the exchange of a series of messages among agents --- conversation.

• The objective of the protocols:– To maximize the payoffs (utilities) of the agents --- in cases where the agents

have conflicting goals or are simply self-interested– To maintain globally coherent performance of the agents without violating

autonomy, i.e., without explicit global control, in cases where the agents have similar goals or common problems, as in distributed problem solving (DPS):• Determine shared goals;• Determine common tasks;• Avoid unnecessary conflicts;• Pool knowledge and evidence.

Agent Interaction Protocols--- Coordination

• Why actions of multiple agents need to be coordinated? because – there are dependencies between agents’ actions, – there is a need to meet global constraints, and – no one agent has sufficient competence, resources or information to

achieve system goals.

• Examples of coordination:– Supplying timely information to other agents– Ensuring the actions are synchronized– Avoiding redundant problem solving

Agent Interaction ProtocolsAND/OR Goal Graph

• The actions of agents in solving goals can be expressed a representation through a classic AND/OR goal graph.

• The goal graph includes a representation of the dependencies between the goals and the resources needed to solve the primitive goals (leaf nodes of the graph). Indirect dependencies can exist between goals through shared resources.

Agent Interaction Protocols

• Coordination protocols• Cooperation protocols• Blackboard systems• Negotiation• Multiagent belief maintenance • Market mechanisms

Agent Interaction Protocols

• Coordination protocols• Cooperation protocols• Blackboard systems• Negotiation• Multiagent belief maintenance • Market mechanisms

Thanks!