Reproduce “Exploration with Multi-Sample Target Values for Distributional Reinforcement Learning” 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
Michael Teng, Michiel van de Panne, Frank Wood
E2DC learns return distributions using multi-sample targets and upper-confidence-bound action selection, extending distributional Soft Actor-Critic. It samples multiple actions to estimate possible future returns, then favors candidates with high predicted means or greater spread. At 200,000 Humanoid-v2 timesteps, it reports reward 2218.3 ± 1394.5, versus IQN’s 1154 ± 249.4. The result is not universal: on Swimmer-v2, its reward is slightly below IQN.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.