Towards Interpretable Policies in Multi-agent Reinforcement Learning Tasks

Deep Learning (DL) allowed the field of Multi-Agent Reinforcement Learning (MARL) to make significant advances, speeding-up the progress in the field. However, agents trained by means of DL in MARL settings have an important drawback: their policies are extremely hard to interpret, not only at the individual agent level, but also (and especially) considering the fact that one has to take into account the interactions across the whole set of agents. In this work, we make a step towards achieving interpretability in MARL tasks. To do that, we present an approach that combines evolutionary computation (i.e., grammatical evolution) and reinforcement learning (Q-learning), which allows us to produce agents that are, at least to some extent, understandable. Moreover, differently from the typically centralized DL-based approaches (and because of the possibility to use a replay buffer), in our method we can easily employ Independent Q-learning to train a team of agents, which facilitates robustness and scalability. By evaluating our approach on the Battlefield task from the MAgent implementation in the PettingZoo library, we observe that the evolved team of agents is able to coordinate its actions in a distributed fashion, solving the task in an effective way.