English

On Convergence of Gradient Expected Sarsa($\lambda$)

Machine Learning 2020-12-15 v1

Abstract

We study the convergence of Expected Sarsa(λ)\mathtt{Expected~Sarsa}(\lambda) with linear function approximation. We show that applying the off-line estimate (multi-step bootstrapping) to Expected Sarsa(λ)\mathtt{Expected~Sarsa}(\lambda) is unstable for off-policy learning. Furthermore, based on convex-concave saddle-point framework, we propose a convergent Gradient Expected Sarsa(λ)\mathtt{Gradient~Expected~Sarsa}(\lambda) (GES(λ)\mathtt{GES}(\lambda)) algorithm. The theoretical analysis shows that our GES(λ)\mathtt{GES}(\lambda) converges to the optimal solution at a linear convergence rate, which is comparable to extensive existing state-of-the-art gradient temporal difference learning algorithms. Furthermore, we develop a Lyapunov function technique to investigate how the step-size influences finite-time performance of GES(λ)\mathtt{GES}(\lambda), such technique of Lyapunov function can be potentially generalized to other GTD algorithms. Finally, we conduct experiments to verify the effectiveness of our GES(λ)\mathtt{GES}(\lambda).

Keywords

Cite

@article{arxiv.2012.07199,
  title  = {On Convergence of Gradient Expected Sarsa($\lambda$)},
  author = {Long Yang and Gang Zheng and Yu Zhang and Qian Zheng and Pengfei Li and Gang Pan},
  journal= {arXiv preprint arXiv:2012.07199},
  year   = {2020}
}

Comments

This submission has been accepted by AAAI2021

R2 v1 2026-06-23T20:56:18.104Z