Reinforcement Learning for Combinatorial Optimization

Abstract

Combinatorial optimization (CO) problems have many important application domains, including social networks, manufacturing, and transportation. However, as an NP-hard problem, the traditional CO problem-solvers require domain knowledge and hand-crafted heuristics. Facing big data challenges, can we solve these challenging problems with a learning structure within a short time? This article will demonstrate how to solve the combinatorial optimization problems with the deep reinforcement learning (DRL) method. Reinforcement learning (RL) is a subfield of machine learning (ML) that learns the optimal policy over time. Building on Markov decision process, RL has the solid theoretical foundation to obtain the optimal solution. Once parameters of DRL are trained, a new problem case can be solved quickly. Moreover, DRL learns the optimal solution without labels by maximizing the accumulative discounted reward received from the environment. This article will discuss three typical CO problems and present the advantages of DRL over other traditional methods.