Bayes-Adaptive Deep Model-Based Policy Optimisation

TL;DR

This paper presents RoMBRL, a Bayesian deep model-based reinforcement learning method that effectively captures model uncertainty for more sample-efficient policy optimization, outperforming existing methods on control benchmarks.

Contribution

Introduction of RoMBRL, a Bayesian deep RL approach using belief distributions and history-based policies, enabling better uncertainty handling and sample efficiency.

Findings

RoMBRL outperforms existing methods on control benchmarks.

The method achieves higher sample efficiency and task performance.

Uncertainty propagation improves policy optimization.

Abstract

We introduce a Bayesian (deep) model-based reinforcement learning method (RoMBRL) that can capture model uncertainty to achieve sample-efficient policy optimisation. We propose to formulate the model-based policy optimisation problem as a Bayes-adaptive Markov decision process (BAMDP). RoMBRL maintains model uncertainty via belief distributions through a deep Bayesian neural network whose samples are generated via stochastic gradient Hamiltonian Monte Carlo. Uncertainty is propagated through simulations controlled by sampled models and history-based policies. As beliefs are encoded in visited histories, we propose a history-based policy network that can be end-to-end trained to generalise across history space and will be trained using recurrent Trust-Region Policy Optimisation. We show that RoMBRL outperforms existing approaches on many challenging control benchmark tasks in terms of sample complexity and task performance. The source code of this paper is also publicly available on https://github.com/thobotics/RoMBRL.