Online Hyper-parameter Tuning in Off-policy Learning via Evolutionary Strategies

TL;DR

This paper introduces an evolutionary strategies-based framework for online hyper-parameter tuning in off-policy learning, addressing the sensitivity of algorithms to hyper-parameters and outperforming existing methods across various benchmarks.

Contribution

It presents a novel application of evolutionary strategies for dynamic hyper-parameter tuning in off-policy learning, bridging connections to meta-gradients and improving performance.

Findings

Outperforms state-of-the-art off-policy algorithms with static hyper-parameters

Effective in continuous control benchmarks

Leverages black-box optimization for low-dimensional search spaces

Abstract

Off-policy learning algorithms have been known to be sensitive to the choice of hyper-parameters. However, unlike near on-policy algorithms for which hyper-parameters could be optimized via e.g. meta-gradients, similar techniques could not be straightforwardly applied to off-policy learning. In this work, we propose a framework which entails the application of Evolutionary Strategies to online hyper-parameter tuning in off-policy learning. Our formulation draws close connections to meta-gradients and leverages the strengths of black-box optimization with relatively low-dimensional search spaces. We show that our method outperforms state-of-the-art off-policy learning baselines with static hyper-parameters and recent prior work over a wide range of continuous control benchmarks.