Imitation Learning as a Tool for Modeling Cooperative Agent Games

The modern day sees constant application of team-level cooperation, and game theory offers a way to model these tasks in a way that provides applicable results. We created a game using Python 3 that had 3 or more agents from one team surround an agent from another team to consume them, with the end goal of consuming the other team. These agents were then trained in a centralized training, decentralized execution schema to mimic an expert’s behavior using imitation learning. Agents trained under this imitation learning schema performed significantly better than agents who followed a random policy, and as more training was supplied to the agents to learn off of, the agents performed better compared to other lesser trained teams.