PLASMAPlanning and Acting in Multi-Agent Systems

The increasing penetration of multi-agent systems in the society will require a paradigm shift—from single-agent to multi-agent planning and reinforcement learning algorithms—leveraging on recent breakthroughs. To this end, this proposal aims at designing a generic software or machine that can efficiently compute rational strategies for a group of cooperating or competing agents in spite of stochasticity and sensing uncertainty, yet using the same algorithmic scheme. Such a machine should adapt to changes in the environment; apply to different tasks; and eventually converge to a rational solution for the task at hand. Our objective is to contribute to theoretical foundations of intelligent agents and multi-agent systems by characterizing the underlying structure of multi-agent decision-making problems and designing efficient planning and reinforcement learning algorithms with performance guarantees. The main idea is that it is possible to reduce a multi-agent decision-making problem (such as a partially observable stochastic game) to a fully observable stochastic game, which is solved using a generic algorithm based on recent advances in Artificial Intelligence.