Reinforcement learning approaches have been applied to replace handcrafted components of meta-heuristics to guide their search processes in a more goal-oriented fashion. While genetic algorithms are well-performing meta-heuristics, it is known that they might suffer from premature convergence due to the loss of diversity in later generations. Therefore, we propose a "Neural Genetic Algorithm" (NGA) that replaces handcrafted survival selection operators with a survival selection policy. Thereby, the survival selection process is modeled as an MDP, whereas the action is defined as an ordered set of solutions, which should survive the current generation. We evaluate the performance of the NGA based on a job shop scheduling problem and represent the solutions as a graph. Therefore, a novel graph-based crossover operation for genetic algorithms is proposed. The results suggest that the NGA can outperform most of the handcrafted survival selection operations significantly, while it struggles to generalize to larger problem instances.      «

Reinforcement learning approaches have been applied to replace handcrafted components of meta-heuristics to guide their search processes in a more goal-oriented fashion. While genetic algorithms are well-performing meta-heuristics, it is known that they might suffer from premature convergence due to the loss of diversity in later generations. Therefore, we propose a "Neural Genetic Algorithm" (NGA) that replaces handcrafted survival selection operators with a survival selection policy. Thereby,...     »