Evolutionary Computation And Its Application In Neural Networks

Evolutionary computation including genetic algorithms, evolutionary progranuning, evolution strategies and genetic programming, is a class of stochastic search algorithms. It has been used to tackle complex questions such as combination optimization, function optimization and machine learning, and has aroused many researchers' concerns and interest.This thesis is focused on the following flve topics.First, comparative analysis of binary encoding and float encoding is made, the advantage and disadvantages of two encoding modes and their influence on genetic operators are clarified, thus the basis for reasonable description of the problems is provided.Secondly, as genetic operators have an important influence on performance of algorithms, this thesis demonstrates that the simulated binary crossover can keep the mean of population unchanged, and under some conditions. can increase the variance of population.The question ho\v to maintain the population diversity has always been one of the main themes of evolutionary algorithms. This thesis experimentally illustrates the effect of adaptive operators and niching method on population diversity.Thirdly, resorting to cooperation-competition model of biomathematics, this thesis proposes a new co-evolution model. Simulation results are shown to verify its effect and practicabilitv.Last, standard methods for optimizing neural netvvorks are easily trapped into local optimization, and unable to adjust the structure of neural networks, thus their application is limited to certain extent. By combining the evolutionary algorithms with artificial neural networks. the simultaneous evolution of architectures and weights of the artificial neural network is conducted, and some meaningful results are achieved.