An Entropy Regularization Free Mechanism for Policy-based Reinforcement Learning
Published in arXiv, 2021
Recommended citation: Changnan Xiao, Haosen Shi, Jiajun Fan, Shihong Deng. "An Entropy Regularization Free Mechanism for Policy-based Reinforcement Learning." arXiv preprint arXiv:2106.00707, 2021. https://arxiv.org/abs/2106.00707
Policy-based RL is prone to policy collapse, which value-based methods avoid through three properties of ε-greedy exploration: closed-form diversity, objective-invariant exploration and adaptive trade-off. This paper proposes an entropy-regularization-free mechanism that gives policy-based methods the same three properties and makes them markedly more sample-efficient on the Arcade Learning Environment.
