Technology

bards-ai/jaxpot

An open-source JAX library for training self-play reinforcement learning agents at scale on the GPU.

Jaxpot is a high-performance Python library designed by bards.ai to accelerate reinforcement learning (RL) training through massive parallelization on a single GPU. Built on top of JAX and optimized for Pgx environments, it enables developers to train AlphaZero-style models, execute Proximal Policy Optimization (PPO), and run league play evaluations at speeds reaching 100 million steps per second. By keeping both the game environments and the agent neural networks entirely on the GPU, Jaxpot eliminates CPU-to-GPU communication bottlenecks: making it ideal for training agents in complex, imperfect-information games like Dark Hex and Liar's Dice.