Novel reinforcement learning methods for routing problems on discrete space, Two-DIMM-per-Channel (2DPC) and PAM-4 interconnection

Abstract

This paper proposes novel reinforcement learning-based routing techniques for various practical tasks. Firstly this research proposes learning $\textit{collaborative policies}$ (LCP), which is a novel hierarchical reinforcement learning method for traveling salesman problem (TSP) and its' variants. Then, this paper extends TSP to hardware routing problems to optimize two-DIMM-per-Channel (2DPC) and PAM-4 interconnection. For hardware routing problems, this paper suggests a novel imitation learning framework. Extensive experiments show our method significantly outperforms baseline reinforcement learning methods.