pgx

Why we included this project

Anyone doing reinforcement learning research on board games has probably felt the pain of environment code that runs one game at a time. pgx sidesteps that by implementing chess, Go, shogi, backgammon, poker variants, and Othello directly in JAX, so the simulators run vectorized across GPU batches instead of looping through Python. That means you can generate thousands of episodes in parallel, which is the kind of throughput AlphaZero-style self-play depends on. The API is consistent across every game, so swapping environments for a benchmark is painless, and there's even an export path to PettingZoo for multi-agent setups. If you're comparing agents across several board games or scaling up self-play runs, this saves you from writing a lot of environment plumbing yourself.