Objective Mismatch in Model-based Reinforcement Learning

TL;DR

This paper identifies a fundamental objective mismatch in model-based reinforcement learning, where optimizing the dynamics model for likelihood does not always improve control performance, highlighting a key limitation and proposing a mitigation approach.

Contribution

The paper characterizes the objective mismatch in MBRL, demonstrating its impact and proposing a re-weighting method to mitigate this issue.

Findings

Likelihood of one-step ahead predictions is not always correlated with control performance

Dynamics models effective for specific tasks may not need to be globally accurate

Global accuracy does not guarantee good control performance

Abstract

Model-based reinforcement learning (MBRL) has been shown to be a powerful framework for data-efficiently learning control of continuous tasks. Recent work in MBRL has mostly focused on using more advanced function approximators and planning schemes, with little development of the general framework. In this paper, we identify a fundamental issue of the standard MBRL framework -- what we call the objective mismatch issue. Objective mismatch arises when one objective is optimized in the hope that a second, often uncorrelated, metric will also be optimized. In the context of MBRL, we characterize the objective mismatch between training the forward dynamics model w.r.t.~the likelihood of the one-step ahead prediction, and the overall goal of improving performance on a downstream control task. For example, this issue can emerge with the realization that dynamics models effective for a specific task do not necessarily need to be globally accurate, and vice versa globally accurate models might not be sufficiently accurate locally to obtain good control performance on a specific task. In our experiments, we study this objective mismatch issue and demonstrate that the likelihood of one-step ahead predictions is not always correlated with control performance. This observation highlights a critical limitation in the MBRL framework which will require further research to be fully understood and addressed. We propose an initial method to mitigate the mismatch issue by re-weighting dynamics model training. Building on it, we conclude with a discussion about other potential directions of research for addressing this issue.