FURI | Spring 2022

Decentralized Multiagent Reinforcement Learning

With the rise of big data, Distributed Computing methods have become the backbone of large systems. Contextualizing problems in a multiagent paradigm enables one to capitalize on the high scalability of such distributed methods. Currently, such multiagent problems are often solved (sub-)optimally using Multiagent Reinforcement Learning (MARL), which relies on a centralized server to perform all the major computation. This work proposes a proof-of-concept that demonstrates the potential of implementing MARL algorithms, namely Multiagent Rollout, in a distributed manner, wherein each agent only uses local information to determine its action which leads to a massive speed-up compared to the centralized approach.