All Tools
M
AgentsFreeOpen Source
MAVA
Distributed multi-agent RL research codebase in JAX
Apache-2.0
ABOUT
Distributed MARL stacks are usually heavy frameworks that are slow to modify. Mava keeps algorithms in single files with Anakin JIT loops for JAX environments and Sebulba for CPU sims, so researchers can hack a baseline and scale it across accelerators without rewriting plumbing.
INTEGRATION GUIDE
1. Clone Mava and train a Sable or other SOTA MARL algorithm on JAX environments
2. Scale an Anakin training loop across GPU or TPU for end-to-end compiled MARL
3. Run Sebulba when the environment is not written in JAX but a GPU learner is available
4. Log JSON metrics that plug into MARL-eval for statistically robust plots
TAGS
pythonjaxmarlmulti-agentreinforcement-learninginstadeep