An improved genetic optimization method for neural network

When GA is used to optimize neural networks, two problems need to be solved. One is the inbreeding and gene coding. Another is the balance between selection pressure and population diversity. One-to-one correspondence between the gene coding and functional equivalence class decreases the coding redundancy through normalizing coding of network. Adaptive crossover and mutation probability is proposed through relating the crossover probability and mutation probability with fitness of individual, which balances the diversity and selection pressure and improves the ability of exploitation and exploration. The simulation experiment shows that the method of optimal design could avoid the inbreeding that results in premature convergence to a certain extent. The method proposed in this paper not only reduces the parameter needed to be learned significantly, but improve the learning efficiency and balance the selection pressure and diversity of population as well.