Predictive model for epistasis-based basis evaluation on pseudo-boolean function using deep neural networks

Complexity of a problem can be substantially reduced through basis change, however, it is not easy to find an appropriate basis in representation because of difficulty of basis evaluation. To address this issue, a method has been proposed to evaluate a basis based on the epistasis that shows the problem difficulty. However, the basis evaluation is very time-consuming. In this study, a method is proposed to evaluate a basis quickly by developing a model that estimates the epistasis from the basis by using deep neural networks. As experimental results of variant-onemax and NK-landscape problems, the epistasis has been estimated successfully by using the proposed method.