@InProceedings{mui-2002c,
  author         = {Lik Mui and Mojdeh Mohtashemi and Ari Halberstadt},
  title          = {Notions of Reputation in Multi-Agent Systems: A Review},
  year           = {2002},
  review-dates   = {2004-08-03},
  value          = {cb},
  address        = {Bologna, Italy},
  publisher      = {ACM},
  month          = {July},
  pages          = {280--287},
  ISBN           = {1-58113-480-0/02/0007},
  hardcopy       = {yes},
  url            = {http://portal.acm.org/citation.cfm?id=544807&dl=ACM&coll=portal},
  organization   = {First International Conference on Autonomous Agents and MAS},
  key            = {mui-2002c}
}

Summary

• a quick review of existing trust and reputation research from reputation reporting systems, economists, scientometrics/bibliometrics, evolutionary biology, and sociologists. There are lots of good references to glean here.
• sythesizes a reputation typology (also found in later in Mui's thesis; mui-2002a) (see below)
• presents a simple reputation simulation based on evolutionary IPD that demonstrates that a strategy's that use more (or better-refined) reputation information dominate those using less (or less-defined) reputation information.

The taxonomy of types presented is as follows:

• reputation
  • individual
    • direct
      • encounter-derived; aka "personalized" in most previous work
      • observed (group)
    • indirect
      • prior-derived
      • group-derived (group)
      • propagated
  • group(group)

But the authors identify contextualization as an independent issue, and claim that all reputation is context dependent. Another split between group vs. personal reputation is identified above in parentheses. [There's a lot to quibble with in this typology]

The paper claims there has been significant confusion between trust and reputation. I believe this paper uses the term reputation in places that trust would better fit. I would claim the category individual.direct would better be called trust, whereas individual.indirect would be called reputation. For instance, an agent's own observations would lend themselves well to the word trust, whereas the center of mass on the definition of the word reputation implies some social standing or some group involved in the assessment.

I'm not sure I follow the thinking behind the individual/group split (as the examples in the group category seem to be just aggregation/binning of individuals, not any novel theory. Though can resolve this by reviewing paper's references in these areas.

I'm not comfortable with all the group/personal quality distinctions.

Other than the sociologist sense (kollock-1994a), I'm not sure the strong distinction is there between encounter-derived and observed reputation (which I would call trust). Furthermore, the authors appear to have conflated the distinction between observations and reported observations in their examples.

OPPORTUNITY: The notion of prior-derived reputation is a good call, and the "role-based trust" of huynh-2004a is an example of this type. However, as this paper says, it hasn't been tried (though outlined in some detail by Huybn). Also notes examples of assumptions in priors of mui-2001a (neutral), and zacharia-1991a (minimal initial reputation).

Follow-ups of interest for me:

• Easy attacks on reputation reporting systems (dellarocas-2000a)
• "Embedded social network" (mui-2002b)
• Bayesian models for reputation (mui-2001a)
• "Chernoff Bounds" on confidence in Bayesian net
• Overlapping groups
• "Chain Store" stage game? (andreoni-1993a, selten-1978a)
• "Folk Theorem"? (fudenberg-1986a)
• Follow up on OTFT (pollock-1992a)
• Social network measure of centrality (katz-1953a, others, see paper)
• Read about personal interactions and affect on trust (kollock-1994a)
• Read on models of group trust (tadelis-2001a, sabater-2001a, halberstadt-2001a)
• "Reciprocity norm" (becker-1990a, gouldner-1960a)

Key Factors

Problem Addressed: Insufficient formalization of reputation types. What is dominance of various types in an evolutionary IPD setting?

Main Claim and Evidence:

• This typology improves on prior efforts. (argumentation primary support here)
• There exists a dominance of [widely propagated reputation] over [locally propagated reputation] over [group reputation] over [observed reputation] over [no reputation model] in an evolutionary IPD game with the insertion of a "reputation TFT" strategy that makes its first move base on

Assumptions:

• Very low repeated encounters between agents is normal for some domains (since the way pairs were linked up were random, in large populations, the encounter-based "reputation" was at a serious disadvantage).
• That group reputation is easily implemented; i.e., that groups will be readily tagged and that AllD's would be distinguishable in the population as part of that group.
• That propagation by the means described (spamming all prior acquaintances by implicit comms seemed to be the way propagation was done in the experiments, rather than incremental gossip between agents about other agents during encounters)

Remaining open questions: What is a really good formal model of trust and reputation ;)?

Quality