Implicit Quantile Networks for Distributional Reinforcement Learning

Abstract

In this work, we build on recent advances in distributional reinforcement learning to give a generally applicable, flexible, and state-of-the-art distributional variant of DQN. We achieve this by using quantile regression to approximate the full quantile function for the state-action return distribution. By reparameterizing a distribution over the sample space, this yields an implicitly defined return distribution and gives rise to a large class of risk-sensitive policies. We demonstrate improved performance on the 57 Atari 2600 games in the ALE, and use our algorithm's implicitly defined distributions to study the effects of risk-sensitive policies in Atari games.

Publication
In International Conference on Machine Learning
Avatar
Will Dabney
Research Scientist

Research scientist trying to invent AI, hopefully doing some science along the way.

Related