1/8 Project DIANE: How Social Structure Improves Distributed Reputation Systems Three Hypotheses Universität Karlsruhe (TH), Germany Institute for Program Structures und Data Organization Universität Karlsruhe (TH) GERMANY Third International Workshop on Agents and Peer-to-Peer Computing (AP2PC 2004) held at AAMAS 2004 Columbia University, New York, USA, July 19th, 2004 Philipp Obreiter, Stefan Fähnrich, Jens Nimis
2/8 agent E agent D Introduction Cooperative behavior in open artificial societies needs incentives Incentive patterns eventually rely on reputation systems P2P networks: reputation system distributed to autonomous agents exchange of recommendations agent C agent Aagent B transaction local instance of reputation system recommendations
3/8 Plausibility Considerations and their Limitations Major challenge: to assess truthfulness of recommendations Existing approaches use plausibility considerations: –Recommendations compatible to first-hand experiences? –How much trust to recommender? (Some) Limitations: –Support for newcomers: newcomers lack support for assessing the recommendations –Recognition of praising: collusions by mutually overstating trustworthiness –Dissemination of recommendations: self-recommendations necessary for effectiveness and efficiency
4/8 Social Structure in Distributed Reputation Systems Social structure: "[…] patterning of interaction, as implying relations between actors or groups, and the continuity of interaction in time." Application to distributed reputation systems: –Relationships: Not necessarily mutual Typed, e.g. (weighted) trust, distrust, bail,… –Dynamics: Not necessarily pre-defined, but adaptive, i.e. relationships can be established and cancelled In our Buddy System: Mutual buddy-relationship: combination of trust- and bail-rel.
5/8 Hypothesis 1: Orientation for Newcomers Hypothesis: Social Structure provides an orientation for newcomers such that they are able to assess the trustworthiness of other agents. Intuition: Newcomers request self-recommendations from agents they are interested in and gain overview of their relationship network Test: Compare performance of newcomers w and w/o social structure Results: Social structure improves performance of newcomers ~20%
6/8 Hypothesis 2: Protection against Collusions Hypothesis: Social structure curbs the impact of colluding agents that mutually praise. Intuition: Explicit collusions become identifiable, recommendations of colluders w/o established relationships become dwarfed Test: Compare performance of colluding and non-colluding agents Results: Effects of praising are marginalized, praising based on relationships is counter-productive
7/8 Hypothesis 3: Effective and Efficient Dissemination Hypothesis: Social structure allows for a more effective and efficient dissemination of recommendations. Intuition: Recommendations disseminated along relationship network are more convincing, agents may self-recommend Test: Show that distributed reputation system with social structure dominates "any other" in terms of effectiveness and efficiency Results: Better efficiency incl. overhead and effectiveness than competitor
8/8 Thank you! More information (especially the technical report) on our project web page: Are there any questions? Thank you for your attention!