Reproduce “Intrinsically-Motivated Reinforcement Learning: A Brief Introduction” 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
Mingqi Yuan
A new intrinsically-motivated reinforcement learning method, Rényi State Entropy Maximization (RISE), significantly enhances exploration efficiency and robustness in complex environments by framing exploration as a Coupon Collector's Problem. It achieves superior average returns and faster learning in tabular, Atari, and continuous control Bullet games, while requiring fewer auxiliary models than previous approaches.
Key insights
- Maximizing Rényi entropy of order in (0, 1) more effectively aligns with the objective of complete state space coverage than Shannon entropy, providing sustained exploration incentives.
- Intrinsic rewards can be efficiently derived from k-NN distances in a learned latent space, mitigating issues like vanishing rewards and the noisy-TV problem without complex auxiliary prediction models.
- Adaptive selection of the 'k' parameter in k-NN estimation is crucial for generating accurate and robust intrinsic rewards across diverse environments and state distributions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.