Trust Analysis through Relationship Identification Ronald Ashri 1, Sarvapali D. Ramchurn 1, Jordi...

33
Trust Analysis through Relationship Identification Ronald Ashri 1 , Sarvapali D. Ramchurn 1 , Jordi Sabater 2, Michael Luck 1 and Nick Jennings 1 1. Intelligence, Agents, Multimedia, University of Southampton 2. Institute of Cognitive Science and Technology, CNR, Roma
  • date post

    19-Dec-2015
  • Category

    Documents

  • view

    219
  • download

    0

Transcript of Trust Analysis through Relationship Identification Ronald Ashri 1, Sarvapali D. Ramchurn 1, Jordi...

Page 1: Trust Analysis through Relationship Identification Ronald Ashri 1, Sarvapali D. Ramchurn 1, Jordi Sabater 2, Michael Luck 1 and Nick Jennings 1 1.Intelligence,

Trust Analysis through Relationship Identification

Ronald Ashri1, Sarvapali D. Ramchurn1, Jordi Sabater2, Michael Luck1 and Nick Jennings1

1. Intelligence, Agents, Multimedia, University of Southampton

2. Institute of Cognitive Science and Technology, CNR, Roma

Page 2: Trust Analysis through Relationship Identification Ronald Ashri 1, Sarvapali D. Ramchurn 1, Jordi Sabater 2, Michael Luck 1 and Nick Jennings 1 1.Intelligence,

Talk Outline

Motivation Relationship Identification Relationship Characterisation Relationship Interpretation

Page 3: Trust Analysis through Relationship Identification Ronald Ashri 1, Sarvapali D. Ramchurn 1, Jordi Sabater 2, Michael Luck 1 and Nick Jennings 1 1.Intelligence,

Motivation (0) Trust

Expectation on the efficiency or effectiveness of an opponent (when it has some opportunity to defect)

Highly context dependent and application specific – hard (or impossible) to design one model for all.

The more information components the better (e.g. Debenham,Sierra,2005, Sabater,Sierra,2002, Huynh et al.,2004, Ramchurn et al, 2004, Patel et al, 2005)

Page 4: Trust Analysis through Relationship Identification Ronald Ashri 1, Sarvapali D. Ramchurn 1, Jordi Sabater 2, Michael Luck 1 and Nick Jennings 1 1.Intelligence,

Motivation Most mechanisms for evaluating trust depend

on using: history of interactions to form Confidence:

recommendations from other agents to get Reputation

1

)),('(),(

n

ConnCon

nw

ConwConwrep

),(),()( 21

Page 5: Trust Analysis through Relationship Identification Ronald Ashri 1, Sarvapali D. Ramchurn 1, Jordi Sabater 2, Michael Luck 1 and Nick Jennings 1 1.Intelligence,

Motivation (2) These face some challenges

Obtaining a history of interactions May take time to build sufficient history to

deduce correctly (may suffer some loss) Which agents to choose first?

Obtaining the recommendations of other agents Assume the recommendations are truthful AND

accurate Which recommendations to give more importance

to?

Page 6: Trust Analysis through Relationship Identification Ronald Ashri 1, Sarvapali D. Ramchurn 1, Jordi Sabater 2, Michael Luck 1 and Nick Jennings 1 1.Intelligence,

Motivation (3) In both of these cases the relationships

between agents are rarely taken into account in manipulating and using the information received

This work provides the foundation for improving trust evaluation by taking into account relationships between agents

Page 7: Trust Analysis through Relationship Identification Ronald Ashri 1, Sarvapali D. Ramchurn 1, Jordi Sabater 2, Michael Luck 1 and Nick Jennings 1 1.Intelligence,

Why take into account relationships?

Relationships can provide more information about the context of interaction

They can reveal whether two agents are in competition, cooperation or inclined to collude

This in turn helps in refining trust evaluations since it provide clues as to how agents may behave

Page 8: Trust Analysis through Relationship Identification Ronald Ashri 1, Sarvapali D. Ramchurn 1, Jordi Sabater 2, Michael Luck 1 and Nick Jennings 1 1.Intelligence,

Approach Relationship Identification

Generic Relationship Identification Model Relationship Characterisation

Application Domain Model Identify of all the possible relationships

which are the most relevant Relationship Interpretation

Use identified relationships and additional context information to derive trust valuations

Page 9: Trust Analysis through Relationship Identification Ronald Ashri 1, Sarvapali D. Ramchurn 1, Jordi Sabater 2, Michael Luck 1 and Nick Jennings 1 1.Intelligence,

Relationship Identification

What are relationships?

Page 10: Trust Analysis through Relationship Identification Ronald Ashri 1, Sarvapali D. Ramchurn 1, Jordi Sabater 2, Michael Luck 1 and Nick Jennings 1 1.Intelligence,

Relationship IdentificationFoundational Concepts (Luck and d’Inverno – SMART)

Attributes are describable features of the environment

An environment is a set of attributes Actions can change environments by

adding or removing attributes A goal is a set of attributes describing

desirable environmental states

Page 11: Trust Analysis through Relationship Identification Ronald Ashri 1, Sarvapali D. Ramchurn 1, Jordi Sabater 2, Michael Luck 1 and Nick Jennings 1 1.Intelligence,

Relationship IdentificationAgents

An agent is described by

Attributes – budget,organisation,products

Actions – selling,buying products

Goals (G) – acquiring information, obtaining a product

Page 12: Trust Analysis through Relationship Identification Ronald Ashri 1, Sarvapali D. Ramchurn 1, Jordi Sabater 2, Michael Luck 1 and Nick Jennings 1 1.Intelligence,

Relationship IdentificationViewable Environment

Agents sense the environment to take decisions about which goals to perform or to verify results of actions

The resulting set of attributes describes a viewable environment (VE)

Page 13: Trust Analysis through Relationship Identification Ronald Ashri 1, Sarvapali D. Ramchurn 1, Jordi Sabater 2, Michael Luck 1 and Nick Jennings 1 1.Intelligence,

Relationship IdentificationRegion of Influence

Agents can affect the environment by performing actions

The set of attributes that they can affect define a region of influence (ROI)

Page 14: Trust Analysis through Relationship Identification Ronald Ashri 1, Sarvapali D. Ramchurn 1, Jordi Sabater 2, Michael Luck 1 and Nick Jennings 1 1.Intelligence,

Relationship IdentificationAgent Interaction Model

Agent A

Environment

viewable environment

region of influence

Page 15: Trust Analysis through Relationship Identification Ronald Ashri 1, Sarvapali D. Ramchurn 1, Jordi Sabater 2, Michael Luck 1 and Nick Jennings 1 1.Intelligence,

Relationship IdentificationAgent Interaction Model

Agent A

Environment

viewable environment

region of influence

Agent B

viewable environment

region of influence

region of influence

Page 16: Trust Analysis through Relationship Identification Ronald Ashri 1, Sarvapali D. Ramchurn 1, Jordi Sabater 2, Michael Luck 1 and Nick Jennings 1 1.Intelligence,

Relationship Characterisation

?

?

Which relationships exist?

?

Page 17: Trust Analysis through Relationship Identification Ronald Ashri 1, Sarvapali D. Ramchurn 1, Jordi Sabater 2, Michael Luck 1 and Nick Jennings 1 1.Intelligence,

Agent-Based Market Model

regimentedBy

SingleMarket

Organisation

Product

AtomicProduct

composedFrom

minCardinality 1

CompositeProduct

requiressells

sellsInaffiliatedTobuysFrom

Agent

consistOf

regimentedBy

Market

regimentedByconsistsOf

CompositeMarket

Institution

consistOf

sells sellsInbuysFromrequires

affiliatedTo

regimentedBy

Page 18: Trust Analysis through Relationship Identification Ronald Ashri 1, Sarvapali D. Ramchurn 1, Jordi Sabater 2, Michael Luck 1 and Nick Jennings 1 1.Intelligence,

Mapping

Buyer A

Environment

market

productto sell

goal(product to buy)

Page 19: Trust Analysis through Relationship Identification Ronald Ashri 1, Sarvapali D. Ramchurn 1, Jordi Sabater 2, Michael Luck 1 and Nick Jennings 1 1.Intelligence,

VEB

Trade-Dep

VEA

ROIA

GB

Page 20: Trust Analysis through Relationship Identification Ronald Ashri 1, Sarvapali D. Ramchurn 1, Jordi Sabater 2, Michael Luck 1 and Nick Jennings 1 1.Intelligence,

VEB

Comp-Sell

VEA

ROIA

ROIB

Page 21: Trust Analysis through Relationship Identification Ronald Ashri 1, Sarvapali D. Ramchurn 1, Jordi Sabater 2, Michael Luck 1 and Nick Jennings 1 1.Intelligence,

VEB

Comp-Buy

VEA

GBA

Page 22: Trust Analysis through Relationship Identification Ronald Ashri 1, Sarvapali D. Ramchurn 1, Jordi Sabater 2, Michael Luck 1 and Nick Jennings 1 1.Intelligence,

VEB

Collaboration

VEA

ROIA

GB

ROIB

GA

Page 23: Trust Analysis through Relationship Identification Ronald Ashri 1, Sarvapali D. Ramchurn 1, Jordi Sabater 2, Michael Luck 1 and Nick Jennings 1 1.Intelligence,

Tripartite Relationships

VEC

VEB

VEA

ROIBGC

ROIAGB

Page 24: Trust Analysis through Relationship Identification Ronald Ashri 1, Sarvapali D. Ramchurn 1, Jordi Sabater 2, Michael Luck 1 and Nick Jennings 1 1.Intelligence,

Relationship Interpretation

Trade-Dep

Competition

Who should I trust??

Coll

Page 25: Trust Analysis through Relationship Identification Ronald Ashri 1, Sarvapali D. Ramchurn 1, Jordi Sabater 2, Michael Luck 1 and Nick Jennings 1 1.Intelligence,

Trust Modelling Confidence:

Direct Interactions Starting value depending on agent’s perception of

environment Reputation:

Witnesses or other interacting agents. Trust function eg.

)()1(),(),( repConT

1

)),('(),(

n

ConnCon

nw

ConwConwrep

),(),()( 21

Page 26: Trust Analysis through Relationship Identification Ronald Ashri 1, Sarvapali D. Ramchurn 1, Jordi Sabater 2, Michael Luck 1 and Nick Jennings 1 1.Intelligence,

Specifying Parameters, how?

Starting confidence

Weights of confidence ratings in the reputation model

Relationships provide a context dependent means of doing this

Page 27: Trust Analysis through Relationship Identification Ronald Ashri 1, Sarvapali D. Ramchurn 1, Jordi Sabater 2, Michael Luck 1 and Nick Jennings 1 1.Intelligence,

Trust Inferences Intensity of Relationships

Socio-Economic concepts Relative value of goods traded (in Trade-Dep or Coll) Relative share of the market (in Comp-Buy, Comp-Sell) Context: C Relationship: R

]1,0[: RCI

Page 28: Trust Analysis through Relationship Identification Ronald Ashri 1, Sarvapali D. Ramchurn 1, Jordi Sabater 2, Michael Luck 1 and Nick Jennings 1 1.Intelligence,

Competition

Give low starting confidence

Give low weights to trust reported by those agents

VEBVEA ROIAROIB

VEBVEAGBA

),(1 RCIsconf

),(1 RCIw

Page 29: Trust Analysis through Relationship Identification Ronald Ashri 1, Sarvapali D. Ramchurn 1, Jordi Sabater 2, Michael Luck 1 and Nick Jennings 1 1.Intelligence,

Collaboration

Start with high confidence (proportional to I(C,R))

Give more weight to reported confidence ratings (Proportional to I(C,R)).

VEBVEA

ROIAGB

ROIBGA

Page 30: Trust Analysis through Relationship Identification Ronald Ashri 1, Sarvapali D. Ramchurn 1, Jordi Sabater 2, Michael Luck 1 and Nick Jennings 1 1.Intelligence,

Dependencies

A depends on B to achieve its goal A will give low starting confidence

B might give high starting confidence (I(C,R)) and may also give more importance to A’s reported trust values (I(C,R)).

VEBVEA

ROIAGB

),(1 RCIsconf

Page 31: Trust Analysis through Relationship Identification Ronald Ashri 1, Sarvapali D. Ramchurn 1, Jordi Sabater 2, Michael Luck 1 and Nick Jennings 1 1.Intelligence,

Collusion B depends on A and B collaborates

with/depends on C.

A will not trust B’s ratings of C if A depends on B and vice versa (decreases with the intensity of B and C’s relationship). E.g.

VEC

VEBVEA

ROIAROIBGC

ROIAGB

),(),(1 21 RCIRCIkw

Page 32: Trust Analysis through Relationship Identification Ronald Ashri 1, Sarvapali D. Ramchurn 1, Jordi Sabater 2, Michael Luck 1 and Nick Jennings 1 1.Intelligence,

Conclusions and Future Work An abstract model to analyse relationships Relationships are important in analysing trust

(e.g. Regret) Can provide agents with a context-dependent

means to define starting confidence and weights

Simulate and evaluate the model with a number of trust metrics

Learn to balance the importance of relationships with that of direct interactions and other information

Page 33: Trust Analysis through Relationship Identification Ronald Ashri 1, Sarvapali D. Ramchurn 1, Jordi Sabater 2, Michael Luck 1 and Nick Jennings 1 1.Intelligence,

Questions?

For more info:-Relationships: R. Ashri and M. Luck, actSMART: Building a SMART system, in Understanding Agent Systems, M. d'Inverno and M. Luck (eds), Springer, 2003

Trust and Reputation Models (Reviews):-S. D. Ramchurn, D. Huynh and N. R. Jennings (2004) "Trust in multiagent systems" The Knowledge Engineering Review 19 (1) 1-25.- Jordi Sabater & Carles Sierra, Review on Computational Trust and Reputation Models, Artificial Intelligence Review, Volume 24, Number 1, September 2005, pp. 33-60(28)