Neural Network Structural Optimization Using Elitist Non-Dominated Sorting Genetic Algorithm (NSGA_Ⅱ)

摘要

This study deals with obtaining optimum parameter for neural network structure in system identification area. The user needs to experiment with a number of architecture parameters such as number of input and output lags, layer, and hidden node. The multiplicity of the model parameters make it difficult to get an optimum structure. Some methods such as trial and error, network growing and pruning have been used to obtain optimum structure, but they are not effective and reliable enough. A multi-objective genetic algorithm is applied by minimizing network complexity and mean square error simultaneously. Results show that the proposed algorithm can successfully model the correct structure of simulated examples. Then the algorithm is applied for real measured data from literature.