Cooperative Multi-Agent Deep Reinforcement Learning in Soccer Domains

摘要

In multi-robot reinforcement learning the goal is to enable a group of robots to learn coordinated behaviors from direct interaction with the environment. Here, we provide a comparison of two main approaches designed for tackling this challenge; namely, independent learners (IL) and joint-action learners (JAL). We evaluate these methods in a multi-robot cooperative and adversarial soccer scenario, called 2 versus 2 free-kick task, with simulated NAO humanoid robots as players. Our findings show that both approaches can achieve satisfying solutions, with JAL outperforming IL.