A Quantitative Trust Model for Negotiating Agents

1
A Quantitative Trust Model for Negotiating
Agents

• Jamal Bentahar, John Jules Ch. Meyer
• Concordia University (Canada)
• Utrecht University (the Netherlands)

Imperial College London, June 08, 2007
2
Overview

• Problem and Motivations
• Negotiation Framework
• Trustworthiness Model
• Implementation
• Related Work and Conclusion

3
Context and Problem

• Multi-agent Systems interacting autonomous
  agents
• Communication Protocols specifying allowed
  communicative acts
• Open and dynamic MAS need flexible protocols
  logic-based dialogue games
• Example negotiation dialogue games
• Security engineering a new challenge in
  agent-based software engineering
• Distributed setting e.g. semantic-grid computing
• Computational efficiency

4
Proposed Approaches for Interacting Agents
Mental Approach
Social Approach
Argumentative Approach
Private states Beliefs, Desires, Intentions, etc.
Public states Social commitments
Argumentation theory reasoning
Allen and Perrault, 1980 Cohen and Levesque,
1990 and others
Singh, 2000 Colombetti, 2000 and others
Amgoud and Maudet, 1999 McBurney et al.,
2002 and others
5
Motivations

• How to trust negotiating agents within a
  multi-agent system
• Resources sharing and mutual access

6
Overview

• ? Problem and Motivations
• Negotiation Framework
• Trustworthiness Model
• Implementation
• Related Work and Conclusion

7
Agent Architecture
8
Negotiation Framework
Specification
Reasoning Semantics
9
Negotiation Framework
Argumentation Theory
10
Dialogue Games

• Abstract structures that can be composed
• Sequencing
• Embedding
• Parallelization
• Argumentation-driven decision making process

Game 1
Game 2
,
//
Game 1
Game 2
11
Dialogue Games Specification

• Initiative / reactive dialogue games
• A simple language
• Cond generating arguments from the agents
  argumentation system

Cond
Action_Ag1
Action_Ag2
12
Agent Communication

• Action_Agi ? Make-Offer, Make-Counter-Offer,
  Withdraw, Satisfy, Violate, Accept, Refuse,
  challenge, Justify, Defend, Attack

13
Argumentation

• The notion of argument
• a pair ltPremises, Conclusiongt
• An argument is a pair (P, c) where P is a set of
  beliefs and c is a formula, such that
• i) P is consistent, ii) P c et iii) P is minimal

14
Argumentation

• Attack relation binary relation between
  arguments
• An argument (P1, c1) attacks another argument
  (P2, c2) iff
• c1 c2 or x P2 c1 x

15
Overview

• ? Problem and Motivations
• ? Negotiation Framework
• Trustworthiness Model
• Implementation
• Related Work and Conclusion

16
Foundation

• Probability function
• Rep A?A?D ? 0, 1
• Local beliefs
• Global beliefs testimonies of witnesses

17
Illustration
18
Assessing Agents Reputation

• Central Limit Theorem and the Law of Large
  Numbers
• If M gt w Then Return True
• Else Return False

19
Timely Relevance Function
20
Reputation Graph

• Algorithm 1 Graph Construction

21
Algorithm2 Node Evaluation
22
Complexity

• Construction of the trust graph with n nodes and
  a edges
• n recursive calls of the function Evaluate-Node
  (Agy)
• Each node is visited once
• Assessing the weight of a node
• Using the weight of its neighbors and input
  edges
• Run time of the reputation algorithm

23
Overview

• ? Problem and Motivations
• ? Negotiation Framework
• ? Trustworthiness Model
• Implementation
• Related Work and Conclusion

24
System Architecture

• The system is designed as a society of
  interacting agents
• Agents are equipped with knowledge bases and
  argumentation systems
• Knowledge bases contain propositional formulae
  and arguments
• Platform Jack Intelligent Agents Java

25
System Architecture
26
Architecture of Negotiating Agent
27
(No Transcript)
28
(No Transcript)
29
Overview

• ? Problem and Motivations
• ? Negotiation Framework
• ? Trustworthiness Model
• ? Implementation
• Related Work and Conclusion

30
Related Work

• Two approach types into trusting multi-agent
  systems centralized and decentralized
• Centralized approaches e.g. eBay and Amazon
  Auctions
• The ratings are stored centrally and summed up to
  give an overall rating
• Reputation is a global single value
• The model can be unreliable, particularly when
  some buyers do not return ratings
• These models are not suitable for applications in
  open MAS such as agent negotiation

31
Related Work

• Three main decentralized approaches
• Building on agents direct experiences of
  interaction partners
• Using information provided by other agents
• Certified information provided by referees

32
Related Work

• Regret
• Direct trust weighted means of all ratings
• Referral
• Direct trust
• Trust network

33
Related Work

• Fire
• Direct interaction trust
• Role-based trust
• Witness reputation
• Certified reputation

34
Conclusion

• Proposition and implementation of a probabilistic
  model to secure negotiating autonomous agents
• Formal and efficient computational framework for
  secure argumentation-based agents in multi-agent
  settings
• Tacking into account the reputation of confidence
  agents
• Considering the timely relevance of the
  transmitted information

35
Future Work

• Reducing the complexity of argumentation-based
  reasoning for agent-oriented systems
• Propositional logic vs. Horn logic
• Evaluate the model using concrete scenarios in
  e-business settings
• A general framework for secure and verifiable
  grid-computing-based applications with the
  underlying formal semantics

36
A Quantitative Trust Model for Negotiating
Agents

• Jamal Bentahar, John Jules Ch. Meyer
• Concordia University (Canada)
• Utrecht University (the Netherlands)

Imperial College London, June 08, 2007