Reproduce “Emphatic Temporal-Difference 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
A. Rupam Mahmood, Huizhen Yu, Martha White, Richard S. Sutton
Emphatic Temporal-Difference (ETD) learning offers a robust method for off-policy value prediction using linear function approximation, achieving guaranteed stability and convergence. This approach integrates user-specified interests and state-dependent parameters to provide flexible control over the learning process, addressing prior divergence issues.
Key insights
- The emphasis mechanism ensures that the expected emphasis weights reflect the occupancy probabilities of the *target* policy, rather than the behavior policy, which is critical for off-policy stability.
- The underlying linear system that ETD(λ) solves has a positive definite matrix, guaranteeing a unique solution and preventing divergence in off-policy learning with linear function approximation.
- Explicitly managing the effective state distribution of updates allows for flexible allocation of function approximation resources according to user-defined interests, independent of the behavior policy's visitation frequency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.