Modeling MultiDimensional Trust

1
Modeling Multi-Dimensional Trust

• Nishit Gujral, David DeAngelis, Karen Fullam,
• K. Suzanne Barber
• May 9th, 2006

2
Goal Driven Behavior

• Autonomous agents can control their own behavior.
• Limited resources force an agent to cooperate
  with a partner to achieve goals.

We propose a domain (problem) independent
mechanism for choosing the best partner agent.
3
Trust

• Interaction in uncertain, open environments
  introduces risk
• Trustworthiness models of potential partner
  agents is necessary to
• Select the most suitable partner
• Avoid risks and maximize reward
• One common method Analyze number of positive and
  negative experiences with a solution provider.
  Jonker and Treur, 1999

4
Single Dimensional Trust
What if the problem changes?
What if no trustworthy provider is available?
S2
S1
Goal-holding Agent
Potential Partner
5
Research Objective
How can an agent choose a partner to maximize its
goal achievement?

• Use multi-dimensional trust to model potential
  partners according to several different metrics

6
Multi-Dimensional Trust

• General trust Built on previous experience, not
  situational trust Marsh, 1994
• Griffiths proposed MD-Trust for the task
  delegation problem Griffiths, 2005
• Here, goal requirements specify the trust
  dimension values.

• Trust is modeled as a composition of constraints
  defined in the context of requestors goal.
• Quality, timeliness, availability

7
Multi-Dimensional Trust
What if the problem changes?
What if no fully trustworthy provider is
available?
S2
Q
T
S1
Q
T
Goal-holding Agent
Potential Partner
8
Why Multi-dimensional Trust?

• Evaluating partners and information sources
  beyond the question, Is source A trustworthy?
  (based on one factor)
• Trust based on a single metric (ie. solution
  quality) is insufficient for accomplishing goals
  with multiple constraints. Maximilien and Singh,
  2005
• Instantly adapt to new goals without training new
  trust models.
• Goal requirements must match partner constraints.

9
Defining Multi-Dimensionality
Expected solution quality
Maps solution quality to goal achievement
Expected completion time
Maps completion time to goal achievement
Domain specified cost
Est. probability partner is available
Expected solution price
10
Partner Selection Algorithm

• Choose the partner with the highest Estimated
  Goal Payoff
• If the selected partner is unavailable, update
  constraint model and choose the next suitable
  partner
• Else, Interact, then update models

Est. Reward Payoff
Failed Interaction Cost
Fixed Interaction Cost
11
Experiment
Demonstrate the advantage of modeling partner
agents using multi-dimensional trust.
Actual Reward Payoff
Unavailability Failure Cost
Interaction Cost
12
Assumptions

• Goal reward functions are domain-based and handed
  to an agent by the system
• Goals are atomic and a single partner can
• accomplish the goal
• The goal-holding agent can choose only one
  partner agent
• Goal-holding agents start off with an optimistic
  point of view

13
Reward Functions
14
Partner Behavior

• Timeliness (tactual)
• Hi v2, Med v5, Lo v10
• Quality (qactual) s0.1
• Hi .66
• Lo 0
• Availability (pactual)
• Hi .66
• 27 Agents, full factorial design

Hidden from Goal-holding agent
15
Partner Constraint Models

• Quality (qest)
• Average quality rating for all previous
  successful interactions
• Timeliness (test)
• Average amount of time needed to provide a
  solution in previous interactions
• Availability (aest)
• percentage of times the partner has been
  available based on all previous invitations to
  interact
• Cost (cp,est)
• Fixed at 5 reward units
• All models are initialized favorably to encourage
  exploration among potential partners

16
Partner Selection Strategies

• Complete
• Consider all available metrics for
  multi-dimensional trust
• Quality
• Consider only the quality of solution that a
  partner agent provides
• Timeliness
• Consider only the duration of time that a partner
  requires to deliver a solution
• Random
• Choose any partner with equal probability

17
Reward Function R1
Performance using reward function R1
18
Reward Function R2
Performance using reward function R2
19
Reward Function R3
Performance using reward function R3
20
Reward Function R4
Performance using reward function R4
21
Conclusions

• A goal-holding agent is capable of accurately
  modeling its potential partners.
• models become more reflective of the partners
  behaviors
• An agent is endowed with the ability to assert
  how much it should trust multiple facets of a
  partners behavior.
• Experimental data verifies that multi-dimensional
  trust improves an agents goal achievement when
  the goal is also multi-dimensional.
• Immediate goal changes can be accommodated
  without rebuilding trust models.

22
Future Work

• Introduce multiple goal-holding agents into the
  same resource constrained environment
• Explore coverage
• Multiple partners may be needed for satisfactory
  goal achievement. Barber and Park, 2004
• Demonstrate the effectiveness of
  multi-dimensional trust in a more realistic
  scenario

23
References

• Barber, K. S. and Park, J., Agent Belief
  Autonomy in Open Multi-Agent Systems, Agents and
  Computational Autonomy, Lecture Notes in
  Computer Science, Springer-Verlag p. 7-16, 2004
• Maximilien, E. M., Singh, M. P., Agent-Based
  Trust Model Involving Multiple Qualities, In
  proceedings of International Conference on
  Autonomous Agents and Multi-Agent Systems
  (AAMAS05), pp. 519-526, The Netherlands, 2005
• Marsh. S. Formalising Trust as a Computational
  Concept. PhD thesis, University of Stirling, 1994

• Jonker C. M. and Treur J. Formal Analysis of
  Models for the Dynamics of Trust Based on
  Experiences, In the Proceedings of The 9th
  European Workshop on Modeling Autonomous Agents
  in a Multi-Agent World Multi-Agent System
  Engineering (MAAMAW-99), pp. 221-231. 1999
• Griffiths, N., Task Delegation using
  Experience-Based Multi- Dimensional Trust, In
  Proceedings of the Fourth International
  Conference on Autonomous Agents and Multi-Agent
  Systems (AAMAS05), pp. 489-496, 2005

24
Experimental Setup Details

• Fixed interaction cost 5
• 27 partner agents
• 1000 simulation cycles
• Runs are averaged over 1000 games

25
Reward function R5
Performance using reward function R5
The algorithm does not perform well in sinusoidal
or periodic reward functions
26

• Drawbacks
• Immediate goal changes require relearning trust
  models
• Multi-faceted goals are difficult to accommodate
• What if no trusted partners are available?