1. Agent Technology for e-Commerce
Chapter 3: Multi-agent Systems
Maria Fasli
Maria Fasli, University of Essex

2. MASs definition
It is rare for an agent to act in isolation and it is even more rare for an agent to be useful on its own
Definition: A multi-agent system (MAS) consists of a network of loosely-coupled computational autonomous agents who can perform actions, they have resources at their disposal and they possess knowledge, capabilities or skills. They are situated in a common environment and they can interact through a set of rules, namely an interaction protocol.

3. Single agent-based systems differ from multi-agent systems:
Environment: agents need to take into account others who may interfere with their plans and goals. They need to coordinate with others in order to avoid conflicts.
Knowledge/expertise/skills: these are distributed
Design: agents need not be homogeneous and may be designed and implemented using different methodologies and languages
Interaction: agents interact following rules of interaction (interaction protocols)

4. Coherence
Coherence, i.e. how well the system behaves as a unit, is important in MASs
Coherence can be measured in different ways depending on the problem domain and system
Usually measured from an external observer’s perspective who ascertains whether or not a system appears to be behaving in a coherent way – not necessarily something that the agents themselves are aware of

5. MASs applications They are particularly suitable for:
Building systems to solve complex problems that cannot be solved by any one agent on its own
Dealing with problems that involve many problem-solving methods, require different types of expertise and knowledge or where there are multiple viewpoints
Creating systems where dynamic reorganization is required
Tasks in which the information resources are distributed

6. MASs advantages
The use of MASs technology offers a number of advantages:
Extensibility and flexibility
Robustness and reliability
Computational efficiency and speed
Development and maintainability
Reusability
Reduced costs

7. MAS challenges
A number of issues with regards to their design and implementation:
Efficient and effective interaction protocols
Task and problem formulation, decomposition, task allocation to individual agents and subtask synthesis
Agent identification, search and location in open systems
Coherent and stable system behaviour while avoiding harmful interactions
Representation of information about the state of the environment, as well as other agents, their actions and knowledge
Team and organization formation
Efficient planning and learning algorithms

8. Closed MASs
Static design with pre-defined components and functionalities
The properties of the system are known in advance:
common language
each agent can be developed as an expert
agents are cooperative
multiple developers can work towards the development of the system at the same time
Example: a MAS in an organization

9. Advantages
Distributed load and expertise
Simplicity and predictability, since
components are known
interaction language and protocols are known
agents usually are cooperative
agents share architecture and software
Disadvantages
Maintenance costs can be high
May be less fault tolerant
Difficult to inter-operate with other systems

10. Open MASs
The system has no prior static design, only single agents within
Agents are not necessarily aware of others – a mechanism for identifying, searching and locating others is required
Agents may be non-cooperative, malicious or not trustworthy
Example: open electronic marketplaces

11. Advantages
Single agent or groups are designed separately (modular)
Flexible and fault tolerant
Evolutionary design
Easier to maintain
Dynamic, open society
Disadvantages
Overall behaviour of the system not predictable
Protocols, languages, ontologies may vary across agents
Malicious behaviour difficult to avoid

12. Interaction
Interactions develop as a result of a series of actions whose consequences influence the future behavior of agents
May be direct or indirect, intended, or unintended
Interaction assumes:
Agents that are capable of acting and/or communicating
Situations that promote interaction
Dynamic elements allowing for local and temporary relationships among agents
Definition: An interaction occurs when two or more agents are brought into a dynamic relationship through a set of reciprocal actions (Ferber 1999)

13. Elements of interactions
Goals: The agents’ objectives and goals are important. Goals of different agents can be conflicting
Resources: agents require resources to achieve their objectives. Resources are not infinite, other agents may want to use them as well. Conflicts may arise
Expertise/skills/capabilities. Agents may lack the necessary skills, expertise of capabilities for accomplishing one or more of their tasks. They may require the ‘help’ of others.

14. Agent characterization
Self-interested/antagonistic agents: have incompatible goals with others and they are interested in maximizing their own utility, not necessarily that of the society as a whole
Cooperative/nonantagonistic agents: have usually compatible goals; they act to maximize their own utility in conjunction with that of the entire system

15. Modes of interaction

16. Interaction protocols
Interaction protocols govern the exchange of a series of messages among agents
Communication protocols provide rules that structure message-passing and produce meaningful dialogues or conversations
Cooperation protocols provide a framework within which agents can coordinate their actions to achieve a complex task or solve a difficult problem in a cooperative way
Negotiation protocols are used in situations where agents have incompatible goals to enable the parties involved to reach a compromise and resolve conflict

17. Agent communication Why is communication important?
Communication is required in MASs where agents have to cooperate, negotiate etc., with other agents
Agents need to communicate in order to understand and be understood by others
Diversity introduces heterogeneity
Can we use natural language as an Agent Communication Language (ACL)? No, natural language is ambiguous

18. Aspects of communication
Communication has three aspects
Syntax: how the symbols of communication are structured
Semantics: what the symbols denote
Pragmatics: how the symbols are interpreted
Meaning: A combination of semantics and pragmatics
An agent can have an active or passive role or even both
An agent can function as a master, a slave or a peer respectively

19. Speech Act Theory
Introduced by Austin (1962) and extended by Searle (1969)
Communication exchange is a form of action: a speech act
An utterance has three aspects
Locution: the physical utterance with context and reference
Illocution: the act of conveying intentions
Perlocution: the action that occurs as a result of the illocution
Identifying the intent or illocutionary force of an utterance is important

20. Illocutionary Force Categories
Assertive: statements of fact
Commissive: commitments
Directive: commands in a master-slave structure
Declarative: statements of fact
Expressive: expressions of emotion

21. Communication protocols
An agent communication language is defined at three levels:
Lower level: specifies method of interconnection
Middle level: specifies format/syntax
Top level: specifies meaning/semantics
Communication protocols need to separate:
semantics of the protocol which must be domain-independent
semantics of the enclosed message which may be domain-dependent

22. ACL components
An agent communication language consists of three components:
An ‘outer’ language (pragmatics): the language used to express the primitives, i.e. the performatives
An ‘inner’ language (syntax): the logical or representation language which is used to write the message itself
Vocabulary (semantics): describes the domain of discourse in terms of concepts and their relationships and prescribes meaning to the terms used

23. Knowledge and Query Manipulation Language
KQML is a high level, message-oriented, communication language and protocol for information exchange
An outer language
There is a separation of protocol semantics from content semantics
All the information required to understand the message is included in the message
KQML is independent of:
the transport mechanism (tcp/ip, corba etc.)
the content language (KIF, SQL etc.)
the ontology assumed by the content

24. KQML message structure
(KQML-performative
:sender <word>
:receiver <word>
:in-reply-to <word>
:language <word>
:ontology <word>
:content <expression> …)
Example
(tell
:sender Agent435
:receiver Agent450
:in-reply-to id-msg-005
:language KIF
:ontology cycbase
:content (salary ))
Reserved keywords for performatives and parameters
The performatives are organized into several categories

25. KQML semantics
Semantics to KQML performatives is given in terms of preconditions, postconditions and completion conditions
Semantics is described in a language which includes operators for expressing an agent’s attitudes: beliefs (BEL), knowledge (KNOW), desires (WANT) and intentions (INT)

26. If A and B are the sender and the receiver respectively
Pre(A) and Pre(B) identify the necessary states for sending and receiving and processing a performative respectively
Post(A) and Post(B) determine the states of the agents after the successful utterance of the performative and its receipt
Completion describes the final state after the conversation has taken place and the original intention for uttering the performative has been accomplished

27. Semantics of tell tell(A,B,X)
Pre(A): BEL(A,X)KNOW(A,WANT(B,KNOW(B,S)))
Pre(B): INT(B,KNOW(B,S))
where S may be any of BEL(B,X) or  BEL(B,X)
Post(A): KNOW(A,KNOW(B,BEL(A,X)))
Post(B): KNOW(B,BEL(A,X))
Completion: KNOW(B,BEL(A,X))

28. Criticisms of KQML KQML has been criticized on a number of fronts:
Some of its performatives (achieve, broker, stream-all) are not really speech acts
Lack of commissive performatives
No security model

29. FIPA ACL
FIPA ACL was part of FIPA’s standardization efforts for interaction protocols
Based on speech acts, messages are communicative acts
It upkeeps the distinction between outer and inner languages and is independent of the inner language
A number of communication exchange protocols based on the FIPA ACL communicative acts have been defined

30. Features of FIPA ACL
All communicative acts can be defined compositionally in terms of two core ones: inform and request
The syntax of performatives is similar to that of KQML
(inform
:sender Agent435
:receiver Agent450
:language Prolog
:ontology companyx
:content “salary( ,32765)”)

31. FIPA ACL semantics
The agents’ states are defined via SL (Semantic Language)
Semantics is based on mental attitudes such as belief, desires and intentions etc.
The meaning of primitives is given in terms of Feasibility Conditions (FPs) and Rational Effects (REs)

32. FP(a) describes the necessary conditions that need to hold true in order for the sender to send communicative act a
RE(a) represents the effect that an agent can expect to occur as a result of performing the act (the purpose of the message) – the RE of a communicative act cannot be guaranteed as the receiver of the message is autonomous and may not conform
<i,inform(j,)>
FP: Bi()Bi(Bifj()Uifj())
RE: Bj()

33. Criticisms of FIPA ACL
Lack of security mechanism (though a proposal had been submitted)
No facilitation primitives

34. Comparing KQML and FIPA ACL
Syntactically similar and both ascribe to the same theory
Semantics differ:
different ways to describe the primitives, i.e. pre, post, completion conditions for KQML, Feasibility Conditions and Rational Effects for FIPA ACL
different language to describe the propositional mental attitudes, e.g. KQML’s BEL operator is not the same as FIPA ACL’s B operator
Different treatment of the ‘administration’ primitives

35. Content Languages
Agents need to be able to represent and reason about:
models of other agents, their beliefs, knowledge, perceptions
tasks, task structure, planning, goals
A common language for representing these things is needed (content language)
Examples of content languages: SQL, Prolog, KIF

36. Knowledge Interchange Format (KIF)
KIF is a content language based on first order logic
Highly expressive and can be used to represent knowledge in declarative form in knowledge bases
Human readable

37. KIF syntax and semantics
KIF is a prefix version of first order predicate calculus
Standard logical operators as well as quantification
(exists ((?x bird) (?y tree)) (on ?x ?y)
One can define objects, relations between objects and rules
(defrelation mother(?x ?y):= (and (female ?x) (parent ?x ?y)))
Knowledge about knowledge can be expressed via the (‘) and (,) operators
(interested agent435 ‘(distance ,?x ,?y ,?z))
Model-theoretic semantics

38. Issues
KIF is probably the most widely used content language for knowledge bases
Highly expressive, but:
complicates the building of fully conforming systems
resulting systems tend to be heavyweight
Its logic-oriented format may not be acceptable to a wider community

39. Dialogues
A series of messages being exchanged comprise a dialogue or conversation

40. A layered model of communication
To enable effective agent communication among agent platforms a layered approach is required
Network Infrastructure
Transport and Signalling (TCP, UDP)
Message Transport Protocol (IIOP, HTTP, RMI)
Message Envelope (FIPA message envelope)
Agent Communication Language (FIPA ACL, KQML)
Content Language (KIF, Prolog, SQL)
Dialogue (Auctions, ContractNet, SNP)

41. Ontologies: The problem
Different agent developers  diversity/heterogeneity of agents
Agents may refer to the same concept under different terms
Client: anyone who buys products
Customer: anyone who buys products
Agents may refer to different concepts under the same term
Client: a computer in the client-server architecture

42. A shared representation is essential for successful communication and interaction between perhaps heterogeneous agents to take place
An ontology is a specification of the objects, concepts, and relationships in an area of interest
An ontology represents a domain from a particular perspective

43. Ontologies are useful:
They define a common vocabulary for those who need to share information
They enable the reuse of domain knowledge
They allow us to encode assumptions about a domain explicitly – these can then easily be changed by simply changing the ontology
They allow us to separate domain from operational knowledge
They can be used by people, software agents, databases and other computer systems

44. Explicit ontologies
Domain knowledge can be encoded implicitly in procedures or explicitly in declarative statements
An explicit ontology is a formal description usually in a logical language of the concepts in a domain of discourse, their properties and relations
Focus is on classes which describe concepts

45. Three aspects
Conceptualization: describes the underlying model of the domain in terms of classes, objects, attributes and relations
Vocabulary: assigns symbols or terms to refer to the classes, objects, attributes and relations
Axiomatization: encodes rules and constraints capturing significant aspects of the domain model

46. Developing ontologies
Ontology development is done in a number of steps:
Determine the domain and the scope of the ontology as well as the purpose for using it
Organize the ontology and design the overall conceptual structure of the domain (pivotal in ontology design and development)
Top-down approach: starts with general concepts and proceeds with specialization
Bottom-up approach: starts with defining specific classes and proceeds with generalization
Mixed

47. Check the ontology for syntactic, logical and semantic inconsistencies
Publish and deploy the ontology
Ontologies need to be maintained, updated and perhaps linked with other ontologies

48. Ontology Web Language
Used to describe the terminology of documents on the web
OWL builds on the Resource Definition Framework (RDF) and proposes a specific vocabulary that provides selected frame and description logic primitives to capture ontologies

49. Three sublanguages with increasing expressiveness:
OWL Lite: mainly intended for class hierarchies and limited constraints. Lowest complexity
OWL DL (includes OWL Lite): intended when completeness and decidability are important
OWL Full (includes OWL DL): maximum expressivity, but no computational guarantees

50. OWL example
An ontology describes classes and relations between classes
Example

51. Cooperative problem-solving
The design of systems of agents that successfully cooperate to solve a difficult and large problem or perform a complex task
Key questions:
How to determine shared goals and common tasks
How to decompose tasks, match subtasks to available agents and synthesize the partial results
How to reconcile different views and avoid unnecessary conflicts
What should be the structure of the system: fixed, self-repairing or self-organizing
How to facilitate convergence on solutions despite incomplete or inconsistent knowledge or data

52. MAS for cooperative problem-solving
Complex Task
Task
Decomposition
Sub-sub-sub tasks performed

53. Task decomposition
Given a complex task, how can this be divided into smaller subtasks?
Can be programmed by the designer during implementation or by the agents using hierarchical planning
Can be done spatially based on the layout of information resources or decision points
Can be done functionally according to the expertise available to the agents

54. Task distribution
Once a complex task has been decomposed the subtasks need to be distributed
Criteria for task distribution
Avoid overloading critical resources
Assign tasks to agents with appropriate capabilities
Make an agent with a wide view assign tasks to other agents
Assign overlapping responsibilities to agents to achieve coherence
Assign closely-related tasks to agents in spatial or semantic proximity
Reassign tasks if necessary for completing urgent tasks

55. Task distribution mechanisms
Market Mechanisms: tasks are matched to agents by generalised agreement
Multi-agent planning: planning agents have the responsibility for task assignment
Organizational structure: agents have fixed responsibilities for particular tasks
ContractNet

56. The ContractNet protocol
Developed by Reid Smith, it is an interaction protocol for co-operative problem-solving among agents
Modelled on the contracting mechanism used by businesses to govern the exchange of goods and services
Provides a solution to the connection problem: finding an appropriate agent to work on a given task

57. Manager
agent
Contractor
Task announcement
Task evaluation
Bid message
Task performed
Reports
Bid evaluation
Award message
Report assimilation
Termination message
Contract established

58. In the original ContractNet implementation each node (agent) comprised of
communication processor
task processor
contract processor
local database
A common internode language with a very simple grammar and a partly domain-independent vocabulary was used for communication

59. Advantages of the ContractNet
Simple to implement, and constitutes the basis for other protocols
Fully distributed
Graceful performance degradation
Applications suitable for the ContractNet protocol are those that feature:
A hierarchy of tasks or levels of data abstraction
Relatively large subtasks which justify considerable effort being put into agent selection
Examples: manufacturing control and scheduling

60. Limitations of the ContractNet
A manager may not receive any bids because:
all potential contractors are busy with other tasks
a potential contractor is idle but may rank the proposed task below other tasks under consideration
no contractors, even if idle, are capable of working on the task (the manager may request immediate responses from contractors with messages such as eligible but busy, ineligible, or uninterested)
Managers are under no obligation to inform potential contractors that a contract has already been awarded

61. An agent may bid on a task for which it is marginally suited and preclude itself from bidding on a much more suitable task announced soon afterwards
Agents have knowledge only of contracts which they are part of. So an agent cannot decide its bids in the light of other existing bids and contracts
Agents do not know how busy other agents are. This is especially important for managers
Agents must know how to decompose tasks/problems and synthesize solutions out of sub-solutions

62. Virtual organizations as MASs
Organizations need to be agile and innovative to be able to withstand competition and respond to the changing needs of the customer force
Virtual organization (VO), virtual enterprise (VE) or agile enterprise: comprises a number of partners that forge a consortium dynamically. Reasons:
Meet customer demand
Offer new or improved services or added value
Exploit a niche in the market
VOs can be naturally seen as MASs
Since agents are autonomous, integration and coordination are fundamental issues

63. Lifecycle of a virtual organization
Creation
One or more agents become aware of the need for a VO
Potential partners are contacted and invited to join
Agent selection is important as it determines future success
Operation
Tasks are distributed to partners who also need to coordinate their actions

64. Maintenance
Essential in a dynamic environment as conditions change
Partners may leave and have to be replaced
Changes in the structure and objectives may be necessary
Self-organization
Disbanding
Marks the end of the VO
When the VO has served its purpose, or is no longer profitable, the services provided are no longer needed or competitive

65. Infrastructure requirements for MASs
To facilitate the creation, operation and interaction of multiple agents an underlying infrastructure is needed
A MAS infrastructure consists of the services, conventions and knowledge and facilitate agent interactions
The MAS infrastructure views agents as black boxes – nothing is known about their internal workings
It is the responsibility of the agent developer to ensure that the agent has appropriate components that allow it to function and interact with other agents and use the services provided by the infrastructure

66. The set of services that are considered essential for the enactment and support of open MASs
Operating Environment
Communication Infrastructure
ACL Infrastructure
Performance and Management Services
Security and Trust Management Services
Agent Description and Discovery Services
Interoperation Services
MAS Infrastructure