NROWAN-DQN: A stable noisy network with noise reduction and online weight adjustment for exploration

作者：

Highlights：

• A differentiable deterministic factor is designed for NoisyNet-DQN.

• A noise reduction method used the differentiable deterministic factor is proposed for NoisyNet-DQN.

• An online weight adjustment mechanism is designed for the noise reduction.

• A series of experiments are conducted to demonstrate the superiority of NROWAN-DQN.

• The stability of NROWAN-DQN in action-sensitive environments is analyzed.

摘要

•A differentiable deterministic factor is designed for NoisyNet-DQN.•A noise reduction method used the differentiable deterministic factor is proposed for NoisyNet-DQN.•An online weight adjustment mechanism is designed for the noise reduction.•A series of experiments are conducted to demonstrate the superiority of NROWAN-DQN.•The stability of NROWAN-DQN in action-sensitive environments is analyzed.

论文关键词：Deep reinforcement learning,Exploration,Noisy networks,Noise reduction,Online weight adjustment

论文评审过程：Received 18 December 2020, Revised 27 December 2021, Accepted 25 April 2022, Available online 6 May 2022, Version of Record 13 May 2022.

论文官网地址：https://doi.org/10.1016/j.eswa.2022.117343