MULTIAGENT SYSTEMS

1
MULTI-AGENT SYSTEMS

• BASIC CONCEPTS

2

• Multi-Agent Systems (MAS)
• (Intelligent Agent Systems)
• Software systems in which program modules are
  given autonomy and intelligence and which
  collaborate to attain system objectives
• MAS are major part of Distributed Artificial
  Intelligence (DAI)
• Fastest growing area in Computer Sciences
• Rapidly being applied to real-life technological
  problems (air traffic control, battle
  simulation, manufacturing, supply chain
  management, etc.)

3

• Multi-Agent Systems are a major component in the
  developing paradigm of Intelligent Systems
• There is a new paradigm emerging which will
  transform society as radically as the emergence
  of steam power and the consequent Industrial
  Revolution. The synergy of advanced
  telecommunications, distributed computing,
  autonomous machines, artificial intelligence,
  virtual reality and other areas is creating and
  shaping this new paradigm of Intelligent Systems.

4

• Intelligent Agents
• An intelligent agent is a software entity which
  exhibits, in some significant measure, autonomy,
  intelligence, and environmental awareness, and
  which interacts with its environment to achieve
  internal goals.
• Co-operation with other agents
• Software Hardware?
• (Agents or Holons or Autonomous XXX?)

5

• Why Agents?
• More usable and understandable structure (e.g.
  spaghetti code vs structure modules).
• (Almost) essential in large distributed systems
  where all subsystems need to be continually
  interchanging information to collectively
  achieve or to maintain some desired state
• (e.g. highly coupled subsystems)
• (cf. human case of multi-disciplinary research
  team)

6
Agent Languages Agent Languages An agent
language is a language for programming software
or hardware agents or agent systems. It should
provide for at least some structural agent
features and may also allow agency attributes to
be directly programmed. It may be incorporated
in a development environment with user-friendly
editors, browsers, etc. which facilitate the
programming. Most existing agent languages are
deliberative (e.g. AGENTO, PLACA, CONCURRENT
METAM, APRIL, MAIL) often related closely (or
less so) to agent theory and logic. Some agent
languages, e.g. ABLE, are reactive languages.
7

• Concurrent Object Languages
• a self-contained concurrently executing object,
  with some internal state that is not directly
  accessible to the outside world, responding to
  messages from other such objects, is very close
  to the concept of an agent
• there are numerous concurrent object languages
  and these have often been used for programming
  agent systems
• systems using concurrent objects involve Actor
  model (Hewitt, 1977 Agha, 1986), ABCL system
  (Yonezawa, 1990), (Briot, 1992)

8

• Shoham (Shoham, 1990, 1993) - Agent-Oriented
  Programming
• a new programming paradigm, based on a societal
  view of computation
• uses the intentional systems approach
• three components
• - a logical system for defining agents mental
  state
• - an interpreted programming language
• - an agentification process, for compiling
  agent programs

9

• third component not yet developed (but might
  follow situated automata paradigm)
• Shohams prototype AOP language is the AGENT0
  system
• logical component of AGENT0 is a logic with
  three modalities belief, commitment and
  ability
• AGENT0 programming language specifies an agent
  as capabilities (what agent can do), initial
  beliefs and commitments, and commitment rules
• each commitment rule has a message condition,
  a mental condition, an action

10

• the message condition is matched against
  received messages and the mental condition is
  matched against beliefs
• if both conditions are satisfied the rule fires
  and agent commits to the rules action
• actions are private (internal execution of
  subroutine) or communicative (message sending)
• three types of message request and
  unrequest to carry out or not perform an
  action, and the inform message for sending
  information
• request/unrequest messages usually cause agents
  commitments to be modified
• inform messages change agents beliefs

11

• Knowledge Representation and Reasoning
• Introduction
• Data is a set of simple descriptions
  information comprises interpretations of data,
  and knowledge is information about information.
• Meta-knowledge is knowledge about knowledge.
• The process of using knowledge to create further
  knowledge is reasoning.

12

• Knowledge Representations
• Numerous ways in which knowledge can be
  represented.
• A fundamental notion in knowledge representation
  is that knowledge comprises object (entities)
  and relationships between these (cf.
  connectionism).
• Different knowledge representations use
  different symbolic representations of the
  objects and the relationships.
• Each such symbolic representation is associated
  with a particular way in which the symbols can
  be manipulated.
• The common types of knowledge representation
  are logic- based semantic networks frames
  object-oriented scripts.

13

• Production Rule System
• Can solve problems using its stored knowledge
  (this essentially is the definition of a KBS).
• Questions the user for additional information,
  where needed.
• Provides answers if the user asks why the system
  asked a question.
• Can provide an explanation of how it derived its
  conclusions.

14

Knowledge Base Rules Facts
Working Memory
SHELL
Inference Engine Inference Control
Knowledge Acquisition Sub-System
Explanation Sub-System
User Interface
Expert Knowledge Engineer
User
Architecture of a Knowledge-Based Expert System
15

• Knowledge Base
• In a production rules system, the
  domain-specific knowledge is stored in the
  knowledge base in the form of rules and facts.
• Each rule consists of a premise and a
  conclusion.
• The premise is often called the condition and
  the conclusion is often called the action.
• The general form of a production rule is
• IF (Premise) THEN (Conclusion)
• e.g.
• IF ((John age 27) and (John not married) and
  (John no kids)
• THEN (John more taxable).
• Each fact in the knowledge base can be regarded
  as a reduced form of a rule e.g.
• THEN (John likes Kate).

16

• Inference Engine
• The two basic functions of the inference engine
  are inferencing and control.
• Inferencing is a procedure (process) by which
  some conclusion/s can be deduced from one or
  more premises, or vice versa
• The commonest basis for inferencing is the rule
  of deductive logic called modus ponens.
• IF A
• THEN B
• IF A THEN B

17
Inference Engine The inference engine works in a
recognize-act cycle 1. It recognizes (by
matching) which rules in the knowledge base can
be satisfied by information in the working memory
(WM) (e.g. such as conclusions recorded as true
from previous cycles). 2. It decides which of
the applicable rules should be used (conflict
resolution). 3. It applies the rule (using
inferencing) and adds information to WM or
deletes an earlier item from WM, then goes back
to step 1. The general strategy embodied in
steps 1 and 2 for selecting eligible rules and
deciding which to apply is part of the control
strategy.
18
Model-Based Architecture (from Jennings, 1993)
19
Alternative Approaches Reactive
Architectures Reactive Architectures A
reactive architecture to be defined as one that
does not include any kind of central symbolic
world model, and does not use complex symbolic
reasoning. These are based on stimulus-response
mechanism. This pattern is easy to see in
Subsumption and in Spreading Activation, but not
in Situated Automata (but it is there in
sophisticated form).
20

• Brooks (Brooks, 1986, 1990, 1991b, 1991a) -
  Behaviour Languages
• behaviour based subsumption architecture
• underlying concepts
• 1. Intelligent behaviour can be generated
  without explicit representations of the kind
  that symbolic AI proposes.
• 2. Intelligent behaviour can be generated
  without explicit abstract reasoning of the
  kind that symbolic AI proposes.
• 3. Intelligence is an emergent property of
  certain complex systems.

21
(No Transcript)
22

• Burmeister et al (Burmeister and Sundermeyer,
  1992 Haddadi, 1994) - COSY
• a hybrid BDI-architecture
• developed for a multi-agent testbed (DASEDIS)
• the architecture has five main components
  sensors, actuators, communications, intention,
  cognition
• sensors are for (non-communicative) perceptual
  input
• actuators carry out (non-communicative) actions
• communications component sends messages

23
(No Transcript)
24

• intention component has long-term goals,
  attitudes, responsibilities and the like the
  control elements taking part in the reasoning
  and decision-making of the cognition component
  (Haddadi, 1994)
• cognition component mediates between intentions
  and beliefs and selects appropriate action/s
• cognition component contains the knowledge base
  of beliefs, and three (procedural) components
• - script execution
• - protocol execution
• - reasoning, deciding, and reacting
• a script is a stereotypical plan for achieving a
  goal

25

• a protocol is a stereotypical dialogue for a
  cooperation framework (e.g. contract net)
• reasoning, deciding and reacting component
• - maintains an agenda containing active scripts
• - a script can be (goal-driven) invoked to
  satisfy an intention or (data-driven) involved
  in response to current situation
• - a filter component which selects scripts to
  execute

26
(No Transcript)
27
What Kinds of Agents? Knowledge Agent Expert in
an area of knowledge (human or artificial) Knowled
ge Server Artificial agent with major
capabilities for storing and retrieving
knowledge (eventually reasoning). Interface
Agent Intelligent assistant.
28
Coach or Tutor Agent Intelligent coach or
tutor. Mediator Agent Coordinates other agents
and resolves conflicts. Knowledge Management
Agent High-level coordination of knowledge
activities for an individual or collaborative
group (e.g. Mediator). Information Search
Agent Travelling searcher, e.g. knowbots,
infobots.
29
Directory Agent Points when queried Where is
XXX? Mentor Agent For higher level expertise or
strategy. Autonomous Agents Hardware
Software e.g. Autonomous vehicles, robots,
etc. Other Agents Citation and document
retrieval Dictionaries Atlases (Geographic
Information Systems) etc. etc.
30
COORDINATION, COOPERATION, NEGOTIATION AND
PLANNING Coordination, Cooperation and
Negotiation are aspects of group or joint
activity of agents. Planning can be an
individual activity of an agent, but in a group
it is normally desirable that the plans be
coherent and contribute towards the achievement
of goals or objectives. The focus in this part
of the course will therefore be how can agents
work together towards the achievement of
objectives.
31
NOTE 1. The following draws significantly from
the text OHare, G.M.P. and Jennings, N.R.,
Foundations of Distributed Artificial
Intelligence, John Wiley, 1996. 2. References
cited in the following sections, unless otherwise
indicated, may be found in that
volume. 3. Unless otherwise indicated, quotations
are from this source.
32

• COORDINATION
• Coordination is a mechanism for ensuring that
  agents activities retain some desired
  relationship/s (sequence, complementarity,
  etc.).
• Control is the extent to which coordination
  information must be implemented by a recipient
  agent.
• The range of control is from none to total.
  Control is inversely related to autonomy (for
  the recipient agent, no control corresponds to
  total autonomy, and being totally controlled
  corresponds to zero autonomy.

33

• The fundamental coordination mechanisms are
• - Mutual adjustment
• Agents share information and resources to
  achieve a common goal, adjusting their
  behaviour according to the behaviour of the
  other agents. No agent has prior control and
  decision making is joint. Coordination in peer
  groups and markets is usually by mutual
  adjustment.
• - Direct supervision
• One agent has some degree of control over
  others and can control information, resources,
  and behaviour. Often established through
  mutual adjustment (e.g. following acceptance
  of employment or a contract).

34
- Coordination by standardization Standard
procedures are established for agents to
follow. In mutual adjustment, are implemented
by acceptance. In direct supervision, are
implemented through (mandatory)
requests. - Mediated coordination A mediator
acts as a facilitator (e.g. finds/routes
information, etc.), a broker (a go-between
and advisor on resource negotiations, etc.),
and a supervisor (exercising some degree of
direct supervision). The first role is
mandatory, but the others are optional. A
mediator facilitates or brokers mutual
adjustment between agents and may also use
direct supervision.
35
- Coordination by reactive behaviour Agents
react to particular stimuli (situations) with
specific behaviours (actions). With
appropriately selected or evolved
stimuli-behaviour groupings and distributions,
system-level patterns of coordinated
behaviour emerge which contribute to the
achievement of common or system goals.
36

• Coordinated Systems
• Mechanisms
• Mutual adjustment operates through peer-to-peer
  interactions Direct supervision uses the
  master-slave approach (in pure case, no
  peer-to-peer interactions allowed) A mediator
  facilitates or brokers mutual adjustment between
  agents and may exercise some degree of
  supervision. Reactive coordination depends on
  appropriate combinations of stimuli-behaviour
  patterns.
• Coordinated Groups and Organization Structure
• Mutual adjustment can work well in small groups
  but the number of information links and amount
  of information increases rapidly with size.

37

• Large groups can be effectively divided into
  sub-groups if most of the information exchange
  can occur in the subgroups.
• Effective coordination of subgroups requires
  each to have a coordinator/supervisor/mediator,
  who work together in one or more coordination
  groups.
• Each coordination group as well as each
  lower-level sub group can be internally
  coordinated either by mutual adjustment, or
  direct supervision, mediation, or reactive
  behaviour.
• A pure hierarchy of the groups results if direct
  supervision is the only coordination mechanism
  used.
• A (pure) heterarchy of agents results if only
  mutual adjustment is used as the mechanism, with
  no groups existing.
• Other intermediate forms of organization result
  from the use of hierarchies/heterarchies a
  hierarchy of mediated groups, etc.

38

• Organization, Information, and Coordination
• An organization is composed of organizational
  entities whose coordination is implemented
  through information flows and exchanges
• Organizational Entities
• - Tasks, parts, machines, tools,
  ----- (what)
• - Functions, processes, ----- (how)
• - Schedules, schedulers, ----- (when)
• - Locations, destinations, ----- (where)
• NOTE All of these organizational entities can be
  represented by agents, in a multi-agent system.

39

• Information
• - Properties, status, -----
• - Conflicts, problems, -----
• - Goals, objectives, -----
• - Structures, membership, -----
• - Relationships, dependencies, -----
• - etc.
• Coordination
• - mutual adjustment
• - direct supervision
• - coordination by standardization
• - mediated coordination
• - reactive coordination

40
NOTE The structure of the organization can be
based on one or more of the organizational
entities, the information infrastructure, or
the coordination mechanisms.
41

• COOPERATION
• Of the five coordination mechanisms considered,
  only two (mutual adjustment, mediated
  coordination) allows for agents to have autonomy,
  i.e. the freedom to make choices and determine
  their own actions. Agents which have some degree
  of autonomy can cooperate with other such agents
  on tasks and activities, for their own or mutual
  benefit.
• Advantages of Cooperation
• Complete tasks quicker through shared effort
• By sharing resources, achieve tasks otherwise
  not possible
• Make use of complementary capabilities
• Avoid harmful interactions

42

• Modes of Cooperation
• accidental not intended
• unilaterally intended one agent intentionally
  helps another
• mutual cooperation two or more agents
  intentionally collaborate
• Degrees of Cooperation
• Fully cooperative (benevolent) Agents always
  attempt to assist other agents that request or
  need their help
• Antagonistic Agents do not cooperate with
  others and may even try to block their goals
• Partly cooperative Agents sometimes or to some
  extent will assist other agents

43

• NEGOTIATION
• The process of improving agreement (reducing
  inconsistency and uncertainty) on common
  viewpoints or plans through the structured
  exchange of relevant knowledge (Durfee, 1989)
• Bargaining (promising something in exchange for
  something else), bidding (offering a service
  or capability at a specific price), and
  contracting (committing to provide a service or
  capability at a specific price) may be part
  of the negotiation process.
• Prior to negotiation an agent needs to
  identify a resource or service it requires but
  cannot supply and then it needs to identify an
  agent/s which it believes potentially could.

44

• Negotiation Agent contacts other agents and
  identifies its need or requests resource or
  service (often under specified conditions)
  other agent/s identify what can be supplied
  under what conditions initiating agent/s
  accepts conditions supplying agent/s commit
  to provide resource or service.
• - In single-stage negotiation, initiating agent
  makes a request and the respondent accepts or
  rejects this.
• - In multi-stage negotiation, agents iterate
  through more than one stage of
  offer/counter-offer.
• - Negotiation protocols a structured procedure
  for one or more stages of the negotiation
  process.

45

• Contract Net Protocol Agents coordinate their
  activities through contracts to accomplish
  specific goals. An agent acting as a manager
  decomposes its contract (the task or problem it
  was assigned with) into subcontracts to be
  accomplished by other potential contractor
  agents. For each subcontract, the manager
  announces a task to the network of agents.
  Agents receive and evaluate the announcement.
  Agents with appropriate resources, expertise,
  and information reply to the manager with bids
  that indicate their ability to achieve the
  announced task. The manager evaluates the bids
  received and awards the task to the most suitable
  agent, called the contractor. Finally, manager
  and contractor exchange information together
  during the accomplishment of the task. (Moulin
  and Chaib-Draa (Course Text)

46

• In Contract Net, a contractor sends status
  (interim) reports, then final report (task
  completion) to manager. The manager may
  terminate a contract prematurely. Idle agents
  may broadcast their availability.
• Basic contract net model does not consider
  interacting tasks (e.g. two agents each need
  all of an available resource several agents
  all need a resource or another agents help at
  or about the same time)
• Possible approaches to interacting tasks
• 1. Agents negotiate directly on resolving
  problem.
• 2. A mediator coordinates the interaction to
  reduce or eliminate the conflict, or
  negotiates with agents on resolving it.

47
3. The agents not only plan how to accomplish
their tasks/goals but communicate these to
others and negotiate potential conflicts,
modifying plans accordingly. NOTE (a) Approac
h 3 is proactive. Approaches 1 and 2 can be
simply reactive ways of dealing with a conflict
problem after it has become evident (e.g.
through conflicting interactions), or can be
negotiation approaches used in 3s pro-active
procedures. (b) Above considers negative
interactions of tasks, i.e. conflicts.
Interactions of tasks can also be positive, by
providing beneficial synergies. While that can
be accidental, it can also be an advantage of
using planning.