Ruggiero Cavallo, David C. Parkes
Imagine a resource allocation scenario in which the interested parties can, at a cost, individually research ways of using the resource to be allocated, potentially increasing the value they would achieve from obtaining it. Each agent has a private model of its research process and obtains a private realization of its improvement in value, if any. From a social perspective it is optimal to coordinate research in a way that strikes the right tradeoff between value and cost, ultimately allocating the resource to one party—thus this is a problem of multi-agent metadeliberation. We provide a reduction of computing the optimal deliberation-allocation policy to computing Gittins indices in multi-armed bandit worlds, and apply a modification of the dynamic-VCG mechanism to yield truthful participation in an ex post equilibrium. Our mechanism achieves equilibrium implementation of the optimal policy even when agents have the capacity to deliberate about other agents' valuations, and thus addresses the problem of strategic deliberation.
Subjects: 7.1 Multi-Agent Systems; 7. Distributed AI
Submitted: Apr 15, 2008