Project
Pokemon TCG Agentic AI
Reinforcement learning pipeline for the Pokemon TCG, a two-player game with hidden information. PPO, behavior cloning, imitation learning, and MCTS; determinization, counterfactual regret minimization, and Bayesian value calibration.
An ongoing research project building reinforcement-learning agents for the Pokemon Trading Card Game — a challenging environment because it is two-player, turn-based, and full of hidden information (prize cards, deck order, the opponent’s hand).
The pipeline combines PPO, behavior cloning and imitation learning with Monte Carlo tree search. To reason about imperfect information, the agent uses determinization, counterfactual regret minimization, and Bayesian value calibration over beliefs about the opponent’s state.