NeuroCrossover: An intelligent genetic locus selection scheme for genetic algorithm using reinforcement learning

Researchers have been studying genetic algorithms (GAs) extensively in recent decades and employing them to address extremely challenging combinatorial optimization problems (COPs). Although GAs achieve superior performance, they are less efficient because most GAs are designed manually without intelligent parameter configuration to support scalable problem-solving strategies and learnable evolutionary operators. To address this issue, machine learning (ML) techniques have been integrated with GAs for operator and parameter selection, however, few studies have focused on intelligent genetic locus selection for influential operators in GAs. To fill this gap, this paper proposes an intelligent genetic locus selection algorithm that serves as the foundation of parameter configuration for critical operators. With the established framework, the Cross Information Synergistic Attention (CISA) model and the n-step proximal policy optimization (PPO) have been utilized to intelligently select the appropriate genetic locus for the most influential phase, i.e., crossover, during the evolutionary process. The proposed NeuroCrossover algorithm is validated on extensive COPs, including the traveling salesman problem, capacitated vehicle routing problem, and bin packing problem. The results demonstrate the efficiency and effectiveness of our algorithm, which outperforms other methods in terms of solution quality, convergence speed, and generalization. For instance, with the CISA, the average percentage gaps of our algorithm are 3.64% and 6.38% for instances in TSPLIB and CVRPLIB, respectively, obtaining gains of about 0.47% and 1.20% compared to those of GA. The proposed algorithm provides a novel solution to lead GAs to an efficient search and improve their scalability and learnability.& COPY; 2023 Elsevier B.V. All rights reserved.