Cooperative Multi-Agent Systems from the Reinforcement Learning Perspective -- Challenges, Algorithms, and an Application

摘要

Reinforcement Learning has established as a framework thatallows an autonomous agent for automatically acquiring -- in atrial and error-based manner -- a behavior policy based on a specification of the desired behavior of the system.In a multi-agent system, however, the decentralization of thecontrol and observation of the system among independent agentshas a significant impact on learning and it complexity.In this survey talk, we briefly review the foundations of single-agent reinforcement learning, point to the merits andchallenges when applied in a multi-agent setting, and illustrateits potential in the context of an application from the fieldof manufacturing control and scheduling.