Opponent-aware, model-based co-training for multi-agent reinforcement learning. The project studies how an agent can keep improving while the other agents in its environment are also learning and changing.
- Multi-agent environments are non-stationary: a policy trained against today's opponents can become stale as those opponents adapt.
- The method co-trains a latent model of the other agents' strategies alongside the policy.
- Planning then uses that learned opponent model instead of treating other agents as fixed environment dynamics.
Collaborative research project; manuscript in preparation. Code and results will be released with the paper.
- Project page: raj-singh.com/projects/model-based-cotraining
- Author: Rajdeep Singh