Analysing emerging algorithms in multi-agent reinforcement learning

Abstract

Reinforcement Learning has witnessed significant advancement in solving various decision-making problems in Machine Learning (ML), most of which involve more than one agent. We categorize such problems as multi-agent problems and utilize Multi-Agent Reinforcement Learn ing (MARL) to solve them. In this project, we shall have a look at a family of multi-agent environments (PettingZoo) and analyze two state-of-the-art multi-agent algorithms - Multi-Agent Deep Deterministic Policy Gradient (MADDPG) and Deep Deterministic Policy Gradient (DDPG). We aim to train the agents in these newly developed multi-agent environments under both algorithms. After that, we shall analyze their training curves using appropriate benchmarking techniques and re-establish how MADDPG’s centralized critic plays an essential role in communication/coordination-based agents.