Diversity-preserving non-destructive operators in genetic programming and their application to breast cancer diagnosis

摘要

In this paper, diversity-preserving non-destructive operators for tree-based genetic programming (GP) are proposed to control code bloat, which is one of the main issues of GP. Firstly, the proposed method is tested using two GP benchmark problems - namely symbolic regression and 11-multiplexer problems. The proposed approach is compared with the traditional standard approach and a crossover hill-climbing approach, which combines non-destructive operators and traditional operators. The newly proposed approach appears superior to the other two compared approaches, in confining intron growth, which is supposed to be the main reason for code bloat, and achieves equal or better performance. When parsimony pressure is applied, the effect of the proposed GP on code bloat is even clearer. The offspring distribution is analysed to illustrate that introns are effectively confined by the new approach. Afterwards these ideas are applied to a real-world problem of breast cancer detection with the Wisconsin Diagnosis Breast Cancer dataset and their ability to solve a real world problem is demonstrated.