Reinforcement learning for dynamic optimization problems

Abstract

Many real-world applications require optimization in dynamic environments where the challenge is to find optima of a time-dependent objective function while tracking them over time. Many evolutionary approaches have been developed to solve Dynamic Optimization Problems (DOPs). However, there is still a need for more efficient methods. Recently, a new interesting trend in dealing with optimization in dynamic environments has emerged toward developing new Reinforcement Learning (RL) algorithms that are expected to give a new breath in DOPs community. In this paper, a new Q-learning RL algorithm is developed to deal with DOPs based on new deifined states and actions that are mainly inspired from Evolutionary Dynamic Optimization (EDO) aiming appropriate exploitation of the strengths of both RL and EDO techniques to handle DOPs. The proposed RL model has been assessed using modified Moving Peaks Benchmark (mMPB) problem. Very competitive results have been obtained and good performance has been achieved compared with other dynamic optimization algorithm.