Pattern Generator for Producing Solid Initial Solutions for Genetic Algorithm

In this study, we apply a model of a Pattern Generator that generates high-dimensional instances based on a given criterion to help solve combinatorial optimization problems. Here, we utilize a supervised neural network to generate initial solutions for Genetic Algorithm (GA). This idea aims to mitigate GA’s difficulty in arriving at good solutions, especially in large problem spaces. The proposed model begins by sampling the problem space and relatively ranks some sample points. The next step is to train a supervised neural network to form hyperplanes that separate the problem space into areas with good initial solutions and regions with poor initial solutions for GA. The neural network is used to iteratively generate instances in the area for good solutions, which are subsequently used to seed the GA. The proposed method ensures that the starting points are better than random seeds, thereby enhancing the GA’s subsequent search process. The proposed method’s efficacy and generalization abilities are tested against 0-1 knapsack problems.