Developing an Integrated Trust and Reputation Model for Open Multi-Agent Systems [Protected Link] Developing an Integrated Trust and Reputation Model for Open Multi-Agent Systems


  author         = {Dong Huynh and Nicholas R. Jennings and Nigel R. Shadbolt},
  title          = {Developing an Integrated Trust and Reputation Model for Open Multi-Agent Systems},
  year           = {2004},
  review-dates   = {2004-08-02},
  value          = {ca},
  note           = {To be published in Knowledge Engineering Review},
  hardcopy       = {yes},
  key            = {huynh-2004a}


This paper presents a scheme called FIRE for handling multiple trust and reputation metrics and combining them. Overall there are a lot of things going on in this paper and its experiments.

The four types of trust presented in the FIRE model.

  1. Interaction trust

    Base trust is recorded in terms between two agents, using a rating database of tuples that has a historic limit. The history is weighted with older dates given less emphasis (the values for the decay of weightings not given in paper). The historic rating tuples are on any number of terms (papers examples are price, quality, delivery). "Interaction trust" (aka direct trust) is handled by a REGRET model of trust.

  2. Role-based trust (novel)

    An at least partially novel, if not particularly well-justified scheme for representing a priori biases to trust is given here. (though presumably these might be adaptable in later work). Rules are given with establish biases towards expecting good or bad results and with what confidence. Combination is additive. Scheme could work for at least some domains, but not sure evidential reasoning would be on firm ground. Paper doesn't really evaluate this aspect of trust, just discusses it and its possibilities.

  3. Witness Reputation

    This is the core indirect trust. Uses form of Yu and Singh's referral system, has branching and finite referral chain length. Appears to be straightforward extension of trust techniques to compiling reputation from witnesses. This area is less well specified formally.

  4. Certified Reputation (novel)

    This area is novel in that an agent can assist another agent in assessing his reputation. The agent to be trusted can tell the agent deciding to trust him what his ratings have been by others. Though this raises many potential difficulties(*), it helps facilitate the transfer of reputation information in some situations that witness reputation could not. (* - agents may not fully disclose all ratings, agents have to trust....without certification... that the ratings being handed over are true, etc.). At the moment this is perhaps "certification free reputation".

Key Factors

How placed in context (other work): Zacharia's SPORAS is used as a system for comparison zacharia-2000a, Ideas of yu-2003a are adapted for network of neighbors, Sabater's REGRET system is used as basis for "interaction trust"; sabater-2003. eBay and Amazon's techniques are discussed.

Problem Addressed: How to improve on witness reputation trust. How to capture "role-based" trust. How to improve trust inference due to multiple aspects presented within paper (interaction, role-based, witness reputation, and "certified reputation") factors and sythesize overall trust.

Main Claim and Evidence: The claims were that certified reputation is novel and helpful and that role-based trust will be useful and can be customized to domain.

Experiments were conducted on a spherical world with producers and consumers of a single service (many similarities to a wireless communications environment) with neighbors defined by distance and service quality degrading by distance (adding more noise/uncertainty into system). Chagnes included: agents moving, agent base performance changes, agent population changes.

Experiments showed that FIRE outperformed SPORAS and SPORAS generally outperformed a no-trust model at the 95% confidence level. As noise and dynamism increased, the trust and reputation models decayed, but FIRE decayed more gracefully than SPORAS. The addition of a certified reputation component to the reputation model outperformed witness reputation alone in this system. Assumptions: Certified reputation trust in paper relied on the truth-telling of agents being rated (attributing other's ratings estimates to themselvs). Unclear how this can be ensured as of this paper.

Next steps: Address means to establish trust in "certified reputation". Also address: learning, dynamic changes to environment, how to set parameters and justify settings

Remaining open questions: Is this certified reputation idea solid or "cerifiable" in the light that it has to trust an agent with his own referral ;). Has or will role-based trust be proven useful in any domains -- or is it just an idea without evaluation? Can the model be effectively extended to handle learning and adaptation to environment changes?


Originality is good.
Contribution/Significance is good.
Quality of Organization outstanding.
Quality of Writing (How hard to extract answer) excellent.