Distributional Reinforcement Learning With Quantile Regression

Abstract

In reinforcement learning (RL), an agent interacts with the environment by taking actions and observing the next state and reward. When sampled probabilistically, these state transitions, rewards, and actions can all induce randomness in the observed long-term return. Traditionally, reinforcement learning algorithms average over this randomness to estimate the value function. In this paper, we build on recent work advocating a distributional approach to reinforcement learning in which the distribution over returns is modeled explicitly instead of only estimating the mean. That is, we examine methods of learning the value distribution instead of the value function. We give results that close a number of gaps between the theoretical and algorithmic results given by Bellemare, Dabney, and Munos (2017). First, we extend existing results to the approximate distribution setting. Second, we present a novel distributional reinforcement learning algorithm consistent with our theoretical formulation. Finally, we evaluate this new algorithm on the Atari 2600 games, observing that it significantly outperforms many of the recent improvements on DQN, including the related distributional algorithm C51.

Asymmetric and adaptive reward coding via normalized reinforcement learning

“…Asymmetric NRL prediction errors also provide a computational mechanism for recent theories of distributional reinforcement learning. In contrast to standard RL approaches, where models learn a single scalar quantity representing the mean reward, distributional RL models learn about the full distribution of possible rewards [33,34]. The critical difference in distributional approaches is a diversity of RPE channels, with differing degrees of optimism about outcomes, which learn varying predictions about future reward.…”

Fusion of Model-free Reinforcement Learning with Microgrid Control: Review and Vision

“…Later, a distributional DQN [56] and a quantile regression DQN [57] were proposed using stochastic policy and distributed training, and they were combined as a 'Rainbow DQN' by David Silver [58] in 2017.…”

Deep Reinforcement Learning Policies Learn Shared Adversarial Features across MDPs

“…The Arcade Learning Environment (ALE) is used as a standard baseline to compare and evaluate new deep reinforcement learning algorithms as they are developed (Hasselt, Guez, and Silver 2016;Mnih et al 2016;Wang et al 2016;Fedus et al 2020;Rowland et al 2019;Mnih et al 2015;Hessel et al 2018;Kapturowski et al 2019;Dabney et al 2018;Dabney, Ostrovski, and Barreto 2021;Xu et al 2020;Schmitt, Hessel, and Simonyan 2020). As a result, any systematic issues with these environments are of critical importance.…”