Multi-Agent Imitation Learning for Agent Typification: A Proof-Of-Concept For Markov Games With Chess
Helen Haase, Daniel Glake, Thomas Clemen · 2024
Cooperative Multi-Agent learning represents a formidable challenge within the realm of artificial intelligence research. This becomes particularly evident in scenarios with high-dimensional environments where agents possess diverse capabilities and must adhere to specific rules, while learning from experts. Imitation learning emerges as a promising solution, allowing agents to acquire policies by studying demonstrations without receiving direct reward signals. We present a proof of concept that applies Multi-Agent Imitation Learning in combination with self-supervised learning using multiple relabeling steps. This method is applied in a Markov game-like environment resembling the game of chess. Heterogeneous agents have been typed, showing their individual capabilities to cooperate for a common goal. Experimental results demonstrate that it is possible to imitate complex behaviors for specific tailored agents. We present a comprehensive exploration of the advantages and limitations in the context of chess, a popular example of a high-dimensional environment with cooperative and adversarial agents.