Deep Variational Reinforcement Learning for POMDPs

TL;DR

This paper introduces deep variational reinforcement learning (DVRL), a method that learns a generative environment model and infers latent states to improve decision-making in partially observable, noisy environments.

Contribution

DVRL combines variational inference with reinforcement learning, enabling joint training of environment modeling and policy optimization for POMDPs.

Findings

DVRL outperforms RNN-based methods on Mountain Hike and flickering Atari tasks.

The method effectively learns latent representations suitable for control.

Joint training improves policy performance in partially observable settings.

Abstract

Many real-world sequential decision making problems are partially observable by nature, and the environment model is typically unknown. Consequently, there is great need for reinforcement learning methods that can tackle such problems given only a stream of incomplete and noisy observations. In this paper, we propose deep variational reinforcement learning (DVRL), which introduces an inductive bias that allows an agent to learn a generative model of the environment and perform inference in that model to effectively aggregate the available information. We develop an n-step approximation to the evidence lower bound (ELBO), allowing the model to be trained jointly with the policy. This ensures that the latent state representation is suitable for the control task. In experiments on Mountain Hike and flickering Atari we show that our method outperforms previous approaches relying on recurrent neural networks to encode the past.