Reinforcement learning enhanced swarm intelligence and trajectory-based algorithms for parallel machine scheduling problems

Abstract

As one of the machine learning methods, reinforcement learning (RL) brings about notable novelties in a wide range of research fields. In several RL strategies, learning is carried out through an agent, a virtual entity that interacts with the environment. It is either rewarded or punished according to the consequences of the actions taken. These mechanisms can be employed as auxiliary procedures in numerous methods, such as metaheuristic algorithms, which are shown to have great potential for RL strategies. In this regard, this study introduces a two-staged approach for parallel machine scheduling problem (PMSP) with release times and sequence-dependent setup times, which has a large number of applications in real-life. The first stage in the proposed approach includes an RL-enhanced Particle Swarm Optimization (PSO) algorithm. The most notable contribution of this stage is that the proposed PSO is capable of self-tuning its parameters and coefficients according to system states. Therefore, any prior parameter setting or calibration is not necessarily required. In the second stage, the best-found solution by PSO is passed to an Iterated Greedy Search (IGS) algorithm, which is a distinguished trajectory-based metaheuristic algorithm. Thus, IGS attempts to further enhance the found initial solution by PSO. All obtained results in the comprehensive experimental study are verified via appropriate statistical tests. The outcomes of this study point out that the adopted RL contributes to the efficiency of the proposed approach for PMSPs. Secondarily, the canonical IGS can be considered a competitive algorithm due to its descent local search procedure.