Multiagent Systems

1
Multiagent Systems

• Shyh-Kang Jeng
• Department of Electrical Engineering/
• Graduate Institute of Communication Engineering
• National Taiwan University

2
Reference

• M. P. Singh, Multiagent Systems A Theoretical
  Framework for Intentions, Know-How, and
  Communications, Springer Verlag, 1994 (Lecture
  Notes in Artificial Intelligence, 799 )
• J. P. Bigus and J. Bigus, Constructing
  Intelligent Agents with Java, 2nd ed., Wiley
  Computer Publishing, 2001
• S. Russell and P. Norvig, Artificial
  Intelligence A Modern Approach, Englewood
  Cliffs, NJ Prentice Hall, 1995

3
Multiagent Systems

• Have multiple intelligent agents working together
  toward a common goal
• Each agent has a limited view of the state of the
  world
• No global control
• Data is decentralized
• Computation is asynchronous

4
Distributed Artificial Intelligence

• Distributed Artificial Intelligence is a modeling
  of artificial intelligence with distributed
  computing
• Artificial intelligence
• Focus on the individual
• Uses psychology as a source of ideas,
  inspiration, and metaphor
• DAI
• Focus on the group
• uses sociology, economics, and management science
  for inspiration

5
Distributed Artificial Intelligence

• DAI also provides insights and understanding
  about interactions among humans, as they organize
  themselves into various groups, committees,
  societies, and economies to improve their lives
• DAI provides a means to construct artificial
  economies that can test economists theories
  before they are applied

6
Distributed Artificial Intelligence

• DAI deals with open systems
• Open systems are subject to new information from
  outside themselves, causing unanticipated
  outcomes
• Most real systems are open system
• DAI is the best way to characterize or design
  distributed computing systems
• DAI provides a natural way to view intelligent
  system
• The environment typically contains other
  intelligent systems

7
Research Directions in DAI

• Extending single-agent concepts to multiple
  agents
• e.g. extensions of belief revision and
  nonmonotonic reasoning to groups of agents
• Developing uniquely multiagent concepts for which
  there are no single-agent analogs
• e.g. negotiation, cooperation, content-based
  communication, multiagent learning, design of
  environments in which autonomous and
  independently-developed agents are guaranteed to
  interact fairly

8
Main Desiderata of Distributed Systems

• Heterogeneous
• Preserve past investment in diverse systems
• Facilitate introduction of new technology
• Optimize platform usage by using the most
  appropriate platform for each task
• Locally Autonomous Components
• Manage security
• Enable incremental change
• Obey legal requirements
• Behave predictably and controllably

9
Intentions, Know-How, and Communications

• Natural to humans
• Provide succinct descriptions of, and help
  understand and explain, the behavior of complex
  systems
• Make available certain regularities and patterns
  of action that are independent of the exact
  physical implementation of the agents in the
  system
• May be used by the agents themselves in reasoning
  about each other

10
Blackboards

• A data structure that is used as the general
  communication mechanism for the multiple
  knowledge sources and is managed and arbitrated
  by a controller
• Each agent looks to the blackboard to pick up new
  information posted by other agents, and it, in
  turn, posts its results to the blackboard
• An event model is used to signal when changes are
  made to the blackboard and to notify the agents
  that something has changed
• Agents are very tightly coupled

11
Communication

• Agents can communicate directly, provided that
  they use the same language, or communicate
  through an interpreter or facilitator
• Two levels of language
• Syntax
• Semantics
• Each specific domain might have its own ontology
• XML gains acceptance in industry and across the
  internet
• In XML a Data Type Definition is used to define a
  particular ontology

12
Knowledge Query and Manipulation Language

• Most widely use agent communication language
  (ACL)
• Came out of the DARPA Knowledge Sharing Effort
• Focuses on message formats and message-handling
  protocols between running agents
• Defines the operations that agents may attempt on
  each others knowledge bases and provides a basic
  architecture for agents to share knowledge and
  information through special agents called
  facilitators

13
Performatives

• KQML messages are called performatives
• Types of speech acts
• Directives
• Representatives
• Commisives
• Every KQML message explicitly states what
  ontology is being used

14
KQML Information Architecture

• Three levels
• Content
• Message
• Communication
• Example
• (ask-one
• sender joe
• content (real price sun.price())
• receiver stock-server
• reply with sun-stock
• language java
• ontology NYSE-TICKS)

Performative
Content level
Communication level
Message level
15
KQML Facilitator

• Two agents who want to communicate using KQML
  require the service of a KQML facilitator
• Agents communicate with the facilitator using
  standard KQML messages
• Agents can register as providers of services or
  information by using the advertise performative
• Agents can ask the facilitator to recommend other
  agents using the recommend, recruit, and broker
  performatives
• The facilitator provides a centralized meeting
  -place for agents and establishes a community
  where agents can interact

16
Agent Standards

• Agent standards are developed to standardize
  agent implementation and to ensure
  interoperability
• Two major efforts
• Foundation for the Intelligent Physical Agents
  (FIPA)
• Focused on agent-level issues
• Object Management Group (OMG) Agent Working Group
• Focus on object-level interoperability and
  management
• MASIF based on aglet system design

17
FIPA Agent Platform
Software
Agent Platform
Agent
Agent Management System
Directory Facilitator
Message Transport System
Message Transport System
Agent Platform
18
Cooperating Agents

• Break a problem into smaller pieces to be handled
  by a collection of smaller and simpler agents
• Will introduce a language barrier between agents
• Examples
• System management
• Electronic commerce
• Collaborative design systems

19
Multiagent Planning

• Extend traditional AI planning algorithms by
  including concepts of joint intentions and
  commitments among agents
• Joint intention means that every team member is
  committed to achieving some goal while also
  believing the other team members are pursuing
  that goal
• Joint commitment is that the team members
  exchange messages to update their mutual beliefs
  and synchronize their behavior
• Centralized planner approach
• Explicit models of teams, teamwork, and team
  interactions

20
Multiagent Planning

• Requires agents share substantial amounts of
  information, incurring significant communication
  and processing overhead
• If the environment changes frequently,
  re-planning is required
• A combination using the team structure and roles
  to limit communications, along with distributed
  planning techniques will provide the best
  solution to building multiagent team

21
Competing Agents

• Everyones agents want to get the best deal for
  their owners
• The winner may be the one with the most
  intelligence, the most specialized knowledge
  about the task it is trying to perform, the most
  powerful reasoning system to apply that knowledge
  to problem solving in the domain, and in all
  likelihood, the one that can learn from
  experience and get better over time, if the
  environment is fair to all agents

22
Contract Net Protocol

• Manager agent calls for bids to complete a task
  (announcement)
• Contractor agents evaluate the announcement and
  optionally respond with their bids
• The manager agent selects one of the contractor
  (bidder) agents
• The contractor agent performs the task and
  returns the results

23
Weakness of Contract Net Protocol

• The contractor agent cannot bid for multiple jobs
  at the same time
• If the contractor got the job receives a better
  offer, the contractor can break his commitment to
  the original manager agent
• This forces another another cycle of bidding
• If an agent is engaged in simultaneous
  conversations with other agents, the bookkeeping
  and overhead can be prohibitive

24
Some Other Protocols

• Auction
• Seller initiates the auction and monitor the
  process while buyers respond with offers to the
  auctioneer
• Has strict rules governing the behavior of the
  auctioneer and the selection of a buyer
• Bargaining
• Proposals and counter proposals may be offered
  back and forth until both parties agree or
  disagree

25
Speech Act

• The action of producing language for
  communication is called speech act
• Forms (e.g. Group of agents in wumpus world)
• Inform Theres a breeze here in 3, 4.
• Query Have you smelled the wumpus anywhere?
• Answer Yes, I smelled the wumpus in 2, 5.
• Request or command Could you help me carry this?
• Promise or offer I will shoot the wumpus.
• Acknowledge OK
• Share That wumpus needs some deodorrant!

26
Formal Language

• Defined as a set of strings
• A string is a sequence of symbols taken from a
  finite set called the terminal symbols
• Most grammars are based on the idea of phrase
  structure
• A string is composed of substrings called
  phrases, which come in different categories
• Phrases are convenient handles on which we can
  attach semantics
• Categorizing phrases helps us to describe the
  allowable strings

27
Sentence

• Sentence (S) is a phrase category representing
  the allowable string
• It can be constructed via combining a noun phrase
  (NP) and a verb phrase (VP)
• Categories like NP, VP, S are nonterminal symbols
• Nonterminal symbols can be used in a grammar rule
  like the rewrite rule
• S ? NP VP

28
Component Steps of Communication

• Speaker S wants to convey proposition P to hearer
  H using words W
• Intention S wants H to believe P
• Generation S chooses W to express P
• Synthesis S utters W
• Perception H perceives W
• Analysis H infers that W has possible meanings
  P1, P2, , Pn
• Disambiguation H infers that S intended to
  convey Pi
• Incorporation H decides to believe Pi

29
Example of Communication Processes
30
Two Models of Communication

• Encoded message
• The speaker has a definite proposition P in the
  mind and encodes the proposition into the words W
• The hearer then tries to decode W to retrieve P
• Situated language
• The meaning of W depends on both the words and
  the situation
• Encoding and decoding functions take an extra
  argument representing current situation

31
Telepathic Communication
ASK(KBB, Q)
TELL( KBA, P)
Agent A
Agent B
Percepts
KB
KB
Percepts
Actions
Actions
Reasoning
Reasoning
32
Problems in Telepathic Communication

• There has to be a naming policy so that A and B
  do not simultaneously introduce the same symbol
  to mean different things
• There has to be some way of relating symbols
  introduced by different agents
• If communication is free and instantaneous, and
  all agents broadcast each new fact to everyone
  else, all agents will have the same knowledge.
  But in most cases, the bandwidth is limited, and
  they are often out of touch. Then they will have
  the problem of deciding what facts are new
• Telepathic communication is vulnerable to
  sabotage, e.g. A tells lies and makes B believes
  anything

33
Communicating using Language
Language
Agent A
Agent B
Percepts
KB
KB
Percepts
Actions
Actions
Reasoning
Reasoning
34
Communicating using Language

• Avoids the sabotage problem
• Infeasible for everyone to have the same internal
  language
• External communication language can be different
  from the internal language
• Agents need not agree on any internal symbols
• Brings the problems of generation and analysis
• Still has problems in reconciling the differences
  between different agents knowledge bases

35
A Typical Dialogue

• ROBOT SLAVE MASTER
• I feel a breeze Go to 1,2
• Nothing is here Go north
• I feel a breeze and I
• smell a stench and
• I see a glitter Grab it

36
A Simple Communicating Agent

• void SimpleCommunicatingAgentrun()
• percept( percept )
• KB.tell(KB.makePerceptSentence(t))
• words percept.SpeechPart()
• semantics
• Disambiguate(Semantics(Parse(words))
• if semantics.type() Command
• action semantics.content()
• else if semantics.type() Statement
• KB.tell(semantics.content())
• action KB.ask(KB.makeActionQuery(percept,t))
  else if semantics.type() None
• action KB.ask(KB.makeActionQuery(percept,t))
  description percept.description()
• Action.Compound(Say(description),Do(action))