Show HN: Play an RL self-play policy in Pokemon Gen 3 Randbats in the browser
Category: ai-ml
Tags: reinforcement-learning, simulation, gpu-computing, game-ai, monte-carlo-tree-search
Score: 7.8/10 (Innovation: 8, Technical: 9, Documentation: 8, Utility: 6)
advsim is a GPU-accelerated, bit-exact reimplementation of Pokémon Showdown's Gen 3 Random Battle simulator using NVIDIA Warp, paired with an RL environment, PPO training, and GPU-based MCTS. It achieves 745K decisions/s on an RTX 4070 and demonstrates a self-play policy reaching 2275 Elo, showcasing a novel combination of high-performance simulation, reinforcement learning, and web deployment.
Target audience: AI/ML researchers, game AI developers, Pokemon enthusiasts
Repository: https://zkingston.com/advsim/ · Python · MIT
View on Hacker News