Methods for optimizing weights of wavelet neural network based on adaptive annealing genetic algorithm

The BP neural network algorithm can not guarantee an error plane as the overall minimum in the training process. It may have a number of local minimum rather than the optimal solution to the issue. To solve this issue, a new genetic algorithm of self-adaptive annealing is designed on the basis of standard genetic algorithm, combined with algorithms for global optimization of simulated annealing to optimize the connection weights. Furthermore, since the standard genetic algorithm is found to have such issues as early immature convergence and late search retardation; difficult coordination of crossover and mutation operator; weak capacity of local search; single way to update the group hard to take care of both the diversity and convergence requirements; slower convergence rate, etc. Based on the characteristics of the algorithm structure, genetic algorithm with parallelism applies the adaptive annealing strategy in calculating the selection probability to enhance the convergence of the genetic algorithm, while adaptive processing is made on the selection of the probability of crossover and mutation to further improve the stability and convergence of the genetic algorithm. The mixed use of genetic algorithms and other algorithms can achieve the advantageous objective while avoiding the disadvantages. Application of this method for training in the Shanghai stock index has witnessed a better network performance.