On the Importance of Hyperparameter Optimization for Model-based Reinforcement Learning

TL;DR

This paper demonstrates that automated hyperparameter optimization significantly enhances the performance of model-based reinforcement learning, reducing the need for expert tuning and improving training stability and rewards.

Contribution

It introduces the use of automatic hyperparameter optimization in MBRL, showing dynamic tuning during training outperforms static hyperparameters and offers insights into hyperparameter effects.

Findings

Automated HPO improves MBRL performance over manual tuning.

Dynamic hyperparameter tuning during training yields better results.

Hyperparameters like plan horizon and learning rate significantly affect training stability.

Abstract

Model-based Reinforcement Learning (MBRL) is a promising framework for learning control in a data-efficient manner. MBRL algorithms can be fairly complex due to the separate dynamics modeling and the subsequent planning algorithm, and as a result, they often possess tens of hyperparameters and architectural choices. For this reason, MBRL typically requires significant human expertise before it can be applied to new problems and domains. To alleviate this problem, we propose to use automatic hyperparameter optimization (HPO). We demonstrate that this problem can be tackled effectively with automated HPO, which we demonstrate to yield significantly improved performance compared to human experts. In addition, we show that tuning of several MBRL hyperparameters dynamically, i.e. during the training itself, further improves the performance compared to using static hyperparameters which are kept fixed for the whole training. Finally, our experiments provide valuable insights into the effects of several hyperparameters, such as plan horizon or learning rate and their influence on the stability of training and resulting rewards.