REGRET: reputation in gregarious societies REGRET: reputation in gregarious societies


  author         = {Jordi Sabater and Carles Sierra},
  title          = {REGRET: reputation in gregarious societies},
  year           = {2001},
  key-cites      = {},
  review-dates   = {2005-02-07},
  editor         = {J{\"o}rg P. M{\"u}ller and Elisabeth Andre and Sandip Sen and Claude Frasson},
  topics         = {},
  address        = {Montreal, Canada},
  pages          = {194--195},
  read-status    = {reviewed},
  hardcopy       = {yes},
  publisher      = {ACM Press},
  citeseer-url   = {"},
  value          = {bb},
  protected      = {no},
  booktitle      = {Proceedings of the Fifth International Conference on Autonomous Agents},
  key            = {sabater-2001a}


The authors claim ReGRET takes each of the three dimensions of trust noted in the paper: individual, social, and ontological. They do in fact present a model that addresses these issues, though the ontological framework is fairly limited. Lots of "linear" weighting parameters used, though some non-linear factors in computing reputation are also used. Many potential moving parts. A great deal of research would be necessary to determine principled means to set the parameters. Dimensions of reputation:

Reputation is defined as the "opinion or view of one about something." [Note: this is a very soft and relativistic definition. Not necessarily any truth ground it.]

Opinions are in part determined by direct interaction. Impressions are recorded as a result of interactions and form the basis of opinions. Opinions are also affected by communication of impressions between agents in the society. Ratings come in range [-1,1] and are subjective. Judgment of ratings should be related to the initial contract, but presumably there is no way to refute a rating... it must be taken at face value.

Addresses reliability by using number of impressions and variability of its ratings values [Similar to SPORAS here].

Stores impressions in an impressions database for i=(a,b,o,\phi,t,W), where a, b are agents, o is an outcome, \phi is the variable of outcome judged, t is the time of the impression, W is a rating.

Pattern for matching IDB of a, IDB^a, as (a,b,o,\phi,t,W):FOL_expr.

When dealing with the social dimension, ReGRET has a factor \omega^{ab_i} that weights how much an agent is like its group. This operates for the personal experience and the group experience.

Then we throw in epsilons to weight the direct interaction, the interaction of a with B's group, the interaction of A's group with B, and A's group with B's group. [How any of the coefficients are set is anyone's idea].

The ontological dimension amounts to a linear combination of various independent judgments of each reputation factor (weighting more important factors more heavily).

The on-line marketplace experiments do not use the ontological dimension, only use the group A to agent b social dimension, and subjective reputation amongst agents is balanced (1/|A|). Reputation reports are made on the agents with quite some precision, even though there is a broad standard deviation, the reports are apparently numeric, and not good/bad. The forget factor is used here. Outperforms SPORAS and Amazon Auctions.

The tourism scenario incorporates the ontological dimension. As they say, the authors "fix a lot of parameters that influence the behavior of the system." [Understandable, given the number of possible parameters to vary.] Here only the direct interactions of travelers with specific travel agencies (agents) and with tour operators (collectives of agents) are considered. Informed travelers (using social dimension) did nearly as well as optimum and much better than those without social dimension or random behavior.

Neither experiment scenario uses the \omega coefficients in reputation of individual agents, elsewhere noted as potentially model ling the "credibility or hierarchical relations".

Key Factors

Relations to Other Work:
  1. Economics: celentani-1966a, marimon-2000a
  2. Psychology: bromley-1993a, karlins-1970a
  3. Computer Science: castelfranchi-1998a, abdul-rahman-2000a, zacharia-1999a, schillo-2000a, yu-2000a
  4. eCommerce: Amazon Auctions, eBay, OnSale Exchange

Problem Addressed: Including components of trust from "different pieces of information". Trust is multi-faceted.

Main Claim and Evidence:

"The paper introduces a new model that takes into account the social dimension of agents and a hierarchical ontology structure."

The authors claim ReGRET takes each of the three dimensions of trust noted in the paper: individual, social, and ontological.


Assumes that Forgetfulness is a good idea.

Appears that ontological dimension's weights are fixed and known up front.

Next Steps:

Extend experiments to validate the model in different situations. Need to extend model to allow for agents to belong to multiple groups at same time. Explore as model for negotiation processes.

Remaining Open Questions:

Is a "forget factor" a generally good idea? How is it tuned? Doesn't it risk confidence games exploiting its forgetfulness?

Is the ad hoc fusion of the number of impressions and variability a good idea? What principled back for this? Any experiments or analysis? How is itm, the tuning/threshold value for absolute reliability w.r.t. number of impressions (it is "application dependent"), determined? Are there cases where high variability in impressions is /not/ an indication of unreliability (or is this an opportunity to identify/look for strategic deception?)

Is the ontological structure methodology (weightings) a good one? Is it just more subjectivity?

Paper hints that ontologies will differ between individuals, but then doesn't show how that plays out or is considered.

Still need to make sense of: "The \omega^{a_ib} and \omega^{a_iB} can reflect the credibility or hierarchical relations inside the group, giving more importance to the information coming from those agents with more credibility or with a higher status in the hierarchy."


Originality is excellent.
Contribution/Significance is good.
Quality of organization is excellent.
Quality of writing is excellent.