Joint Equilibrium Policy Search for Multi-Agent Scheduling Problems

We propose joint equilibrium policy search as a multi-agent learning algorithm for decentralized Markov decision processes with changing action sets. In its basic form, it relies on stochastic agent-specific policies parameterized by probability distributions defined for every state as well as on a heuristic that tells whether a joint equilibrium could be obtained. We also suggest an extended version where each agent employs a global policy parameterization which renders the approach applicable to larger-scale problems. Joint-equilibrium policy search is well suited for production planning, traffic control, and other application problems. In support of this, we apply our algorithms to a number of challenging scheduling benchmark problems, finding that solutions of very high quality can be obtained.