A Mechanism of Generating Joint Plans for Self-interested Agents, and by the Agents

Abstract

Generating joint plans for multiple self-interested agents is one of the most challenging problems in AI, since complications arise when each agent brings into a multi-agent system its personal abilities and utilities. Some fully centralized approaches (which require agents to fully reveal their private information) have been proposed for the plan synthesis problem in the literature. However, in the real world, private information exists widely, and it is unacceptable for a self-interested agent to reveal its private information. In this paper, we define a class of multi-agent planning problems, in which self-interested agents' values are private information, and the agents are ready to cooperate with each other in order to cost efficiently achieve their individual goals. We further propose a semi-distributed mechanism to deal with this kind of problems. In this mechanism, the involved agents will bargain with each other to reach an agreement, and do not need to reveal their private information. We show that this agreement is a possible joint plan which is Pareto optimal and entails minimal concessions.