Heuristic Strategy and Incentives in Distribution Networks of Competitive Agents STIET Fall Workshop, 2003 Presenter: Thede Loder
The Problem Agents (peers) provide or receive services Goal is to acquire valuable objects (assets) Agents have varying taste in assets Agent resources are constrained Free for all environment: –No contracts, no numeraire, no centralized or trusted entities Decision problem: what services, to which agents, and when?
Asset Exchange as a Game Roles: Provider and Initiator Game: iterative, half-step Prisoner’s Dilemma Agent controlled finite horizon (either agent may leave game at any time) Provider takes risk: initiator can get payoff and leave Nash Equilibrium: provide nothing
Heueristic Insight: Axelrod and the Evolution of Cooperation Context: Iterated Prisoner’s Dilemma Tit for Tat (with bluffing): –“exploitive cooperation”, but still cooperation Key insight: cooperation requires expectation of future interaction Bottom line: –No identity No trust No cooperation
Solution Reputation System –Add Identity (cheap by strong) –Add History (agent gets “memory”) Determine reputation by direct and shared experience Make sticky with incentives: differentiated service
Credibility Ratings Simple online learning Use relationship history (actions, outcomes) and value weights Two ratings: –Desirability, influences proactive actions –Credibility, influences reactive actions Ratings effect actuator operation: –Prioritized service queues –introductions and referrals –proactive connection initiation (for network clustering)
A Candidate Function Properties (Axelrod): –Nice, provokable, forgiving Credibility C for an agent is function of: –Net Exchange Volume V n = events * value/event –Correction Factor f(b), function of flow bias b (relative direction of flow) C = V n * f(b) 100% out 100% in symmetric f(b) flow bias
Progress and Next Steps Progress: –Simulator kernel completed –Dummy Agent completed Next Steps: –Parameterize Strategy –Develop ‘smart’ agent code –Run evolutionary search trials