A Study On Theory And Application Of Neural Networks Based Hybrid Intelligence Learning Algorithms

Artificial neural networks and genetic algorithm are the two major branches of artificial intelligence. In recent years, not only the related theories but also the applications have made an encouraging improvement. Considering the characters of BP neural network, such as the simple structure, the advisable malleability, self-fitness, self-studying, nonlinear function approximating, the considerable abilities of parallel computing, fault-tolerant and so on, the BP algorithm have been extensively applied to the areas of system modeling, pattern recognition and seismic exploration since 1986.Compared with other algorithms, as the above reasons, the BP algorithm has become the most usual and efficient solutions to the artificial neural networks.Firstly, this paper analyzes the development and the current situation of the neural networks and genetic algorithm, the related theories of genetic algorithm, such as the basic concept, components, learning rule and simple genetic algorithm, and applies genetic algorithm to the problem optimizing the connection weight of the feedforward neural networks. Secondly, this paper has researched the related theories of nerve cell, characters and learning rule of artificial neural networks, the related theories and performance of BP neural networks. Thirdly, considering the characters of BP neural networks which is good at local minimum and bad in global optimization and the feature of GA neural networks which is bad in local minimum and good at global optimization, the paper proposes a new algorithm combined GA with BP, referred as to hybrid intelligence learning algorithm, which is applied to the problem optimizing the connection weight of the feedforward neural networks. The algorithm has the virtue of neural networks, genetic algorithms and BP algorithms, whose feasibility and utility have been proved in theory and practice. Finally, the paper has designed the program of BP neural networks, neural networks based genetic algorithms and hybrid intelligence learning algorithms in VC++, and applied those algorithms to the XOR problem, the function approximating problem and the explaining high difference seismic data problem. The experiment results have showed that hybrid intelligence learning algorithm for training neural networks is better, faster and more accurate than BP algorithm and genetic algorithm.