One Step at a Time: Pros and Cons of Multi-Step Meta-Gradient Reinforcement Learning

TL;DR

This paper investigates the use of multi-step meta-gradient reinforcement learning, revealing increased variance issues and proposing a mixing method that balances bias and variance, improving stability and performance.

Contribution

It introduces a novel mixing meta-gradient method that reduces variance and enhances robustness in multi-step meta-gradient reinforcement learning.

Findings

Multi-step meta-gradients increase learning signal but also variance.

The mixing method reduces variance by a factor of 3.

The proposed approach achieves comparable or better performance on the Snake game.

Abstract

Self-tuning algorithms that adapt the learning process online encourage more effective and robust learning. Among all the methods available, meta-gradients have emerged as a promising approach. They leverage the differentiability of the learning rule with respect to some hyper-parameters to adapt them in an online fashion. Although meta-gradients can be accumulated over multiple learning steps to avoid myopic updates, this is rarely used in practice. In this work, we demonstrate that whilst multi-step meta-gradients do provide a better learning signal in expectation, this comes at the cost of a significant increase in variance, hindering performance. In the light of this analysis, we introduce a novel method mixing multiple inner steps that enjoys a more accurate and robust meta-gradient signal, essentially trading off bias and variance in meta-gradient estimation. When applied to the Snake game, the mixing meta-gradient algorithm can cut the variance by a factor of 3 while achieving similar or higher performance.