DEEP REINFORCEMENT LEARNING IN MULTI-AGENT PATH PLANNING

Abstract

n this thesis, we introduce the basic concepts of decentralized RL and discuss how they can be applied to MAPP. Our proposed framework consists of only a decentralized decision-making module, where each agent learns its own policy without any centralized coordination. The key advantage of decentralized RL is its ability to handle an increasing number of agents without imposing an excessive computational burden. This scalability ensures that the system’s performance remains unaffected even as the number of agents involved grows. We also assume that communications are available within the team to share relevant information in MAPP.