Deep reinforcement learning with credit assignment for combinatorial optimization

作者：

Highlights：

• Deep Reinforcement Learning is efficient in solving some combinatorial optimization problems.

• Credit assignment can be used to reduce the high sample complexity of Deep Reinforcement Learning algorithms.

• Model-free and model-based reinforcement learning algorithms can be connected to solve large-scale problems.

• Assign credits for hundreds of thousands of state-action pairs in a systemic manner will accelerate the training process.

摘要

•Deep Reinforcement Learning is efficient in solving some combinatorial optimization problems.•Credit assignment can be used to reduce the high sample complexity of Deep Reinforcement Learning algorithms.•Model-free and model-based reinforcement learning algorithms can be connected to solve large-scale problems.•Assign credits for hundreds of thousands of state-action pairs in a systemic manner will accelerate the training process.

论文关键词：Combinatorial optimization,Reinforcement learning,Credit assignment

论文评审过程：Received 30 March 2021, Revised 25 November 2021, Accepted 26 November 2021, Available online 27 November 2021, Version of Record 13 December 2021.

论文官网地址：https://doi.org/10.1016/j.patcog.2021.108466