Neural Evolution Based On Immune Evolutional Algorithm

In the last few years, we could perceive a great increase in studying biologically inspired system. Among these, neural networks, evolutionary computation, DNA computation, and immune system are paid attention to especially. Immune neural evolution is a complex of immune algorithm, evolutionary algorithm and neural networks which has proven to be a capable of performing several tasks, such as pattern recognition, learning, memory acquisition, generation of diversity, noise tolerance, generalization, distributor detection and optimization. Based on biological principles, new computation techniques are being developed, aiming not only at a better understanding of the system, but also at solving engineering problem.Referring to the concept and theory of Immunology in biotic science, the theory, algorithm and application of immune neural evolution were done research on in this paper. First, a simple introduction of immune algorithm, evolutionary algorithm and neural network was offered and the feasibility of combining them was analyzed. Then two algorithms including immune evolution strategy and immune genetic algorithm were proposed. And last, the two proposed algorithms were applied to neural evolution. In the neural evolution based on immune evolution strategy, the antibody encoding, affinity functions design, thickness function design, selection based both on affinity and thickness and the mutation operation were offered. In the immune evolution strategy, the capability of the mutation operator decides the global convergence of the algorithm, so the author did researches on different distributions and proposed a Cauchy mutation operator whose simulation results shows better than the traditional Gauss mutation operator. Simulation results shows that the neural evolution based on immune evolution strategy can meet the global convergence level while the convergent speed may be a little slower because of the encoding. So in the neural evolution based on immune genetic algorithm, an improved coding method coupled with a set of affinity function, thickness function, cross operator and mutation operator were proposed. Simulation experiments compared the two neural evolution methods and the result showed that the neural evolution based on immune genetic algorithm behaved better convergent speed than the neural evolution based on immune evolution strategy. And the two methods show almost the same global convergence capability. These two methods get optimal net structure with optimal weights.