An improved evolution algorithm using population competition genetic algorithm and self-correction BP neural network based on fitness landscape

Abstract

The genetic algorithm-backpropagation neural network algorithm (GA-BP) makes full use of the advantages of genetic algorithm (GA) and BP neural network (BPNN). It has been widely used in practical problems, but it also has shortcomings such as slow convergence. Inspired by the generative adversarial network, a population competition mechanism (PCM) is proposed to improve the search ability of the GA, two populations are generated to compete, and the winning population can obtain the optimal individual from the failed population to ensure that the winning population has a sustainable advantage. The failed population will randomly generate new individuals and add new genes so as to get better individuals, through such a mechanism to ensure the rapid optimization of the entire population, avoid the risk of premature convergence, speed up the iterative speed and improve the stability of the GA. According to the characteristics of the fitness landscape, the learning rate of the BPNN is optimized, and it can change adaptively, which can effectively improve the network convergence speed and greatly reduce the time cost. We define the GA-BP that improves the learning rate based on fitness landscape as FL-GA-BP algorithm. On the basis of the FL-GA-BP algorithm, adding GA improved with PCM, we define this new algorithm as I-GA-BP algorithm, namely the I-GA-BP algorithm that combines PCM-improved GA and fitness landscape-improved BPNN. In this paper, we use two types of test functions with different characteristics and complexity to conduct experiments to verify the effectiveness of the I-GA-BP algorithm. By comparing the experimental data of the three algorithms GA-BP, FL-GA-BP and I-GA-BP, it is obtained that the I-GA-BP algorithm can better escape from the local optimal solution, which is more conducive to finding the global optimal solution. It also greatly improves the convergence speed of the neural network. Finally, we briefly discussed the effect of adjusting the number of neurons on the stability of the I-GA-BP algorithm.