✳flâneur — a map of the web's best reading

JaxMARL: Multi-Agent RL, but 10000x Faster

blog.foersterlab.com · 2,333 words · saved by 1 readers

We present JaxMARL, a library of multi-agent reinforcement learning (MARL) environments and algorithms based on end-to-end GPU acceleration that achieves up to 12500x speedups. The environments in JaxMARL span cooperative, competitive, and mixed games; discrete and continuous state and action spaces; and zero-shot and CTDE settings. We specifically include implementations of the Hanabi Learning Environment, Overcooked, Multi-Agent Brax, MPE, Switch Riddle, Coin Game, and Spatial-Temporal Representations of Matrix Games (STORM). Because of JAX's hardware acceleration, our per-run training pipeline is 12500x faster than existing approaches. We also introduce SMAX, a vectorised version of the popular StarCraft Multi-Agent Challenge, which removes the need to run the StarCraft II game engine. By significantly speeding up training, JaxMARL enables new research into areas such as multi-agent meta-learning, as well as significantly easing and improving evaluation in MARL. Try it out here: htt

JaxMARL: Multi-Agent RL, but 10000x Faster Overview We present JaxMARL , a library of multi-agent reinforcement learning (MARL) environments and algorithms based on end-to-end GPU acceleration that achieves up to 12500x speedups. The environments in JaxMARL span cooperative, competitive, and mixed games; discrete and continuous state and action spaces; and zero-shot and CTDE settings. We specifically include implementations of the Hanabi Learning Environment, Overcooked, Multi-Agent Brax, MPE, Switch Riddle, Coin Game, and Spatial-Temporal Representations of Matrix Games (STORM). Because of JA

Explore this link on the map →

related reading