Reproduce “Jumanji: a Diverse Suite of Scalable Reinforcement Learning Environments in JAX” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Clément Bonnet, Daniel Luo, Donal Byrne, Shikha Surana, Sasha Abramowitz, Paul Duckworth, Vincent Coyette, Laurence I. Midgley and 16 more

InstaDeep researchers present JUMANJI, a JAX-based suite of 22 reinforcement learning environments designed for speed, flexibility, and scalability, with a strong focus on industrial combinatorial optimization problems. The framework demonstrates linear scaling of A2C training speed with increased TPU cores and enhances agent generalization through diverse problem instance generation.

Key insights

  • JAX's composable transformations and stateless environment design enable significant computational speedups and efficient parallel scaling of RL training on hardware accelerators, accelerating research iteration.
  • Training RL agents on diverse problem instance distributions, facilitated by JUMANJI's generator mechanism, can lead to improved generalization capabilities on unseen, real-world problems.
  • Industrial-inspired NP-hard combinatorial optimization problems remain challenging for current RL algorithms, as indicated by persistent optimality gaps even in hardware-accelerated environments, highlighting areas for future research.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.