A Modified Differential Evolution Algorithm and its Application on Neural Network Training

DE has excellent global convergence properties, but its local search ability and search efficiency are limited. A modified differential evolution(MDE) algorithm with hybrid mutation strategy was proposed, its main idea is to make some individuals in the population search around the current best individual, and the other individuals search randomly with its basic mutation strategy, benchmark results show that the search efficiency has been improved above 30%. Then the procedure and method of neural network training with MDE were described, the performance of MDE on neural network training was compared with PSO DE and BP algorithm. Test results show that MDE generally outperforms the other algorithms.