A Combination of Genetic Algorithms and Steepest Descent Method to Improve the Learning Performance of Neural Network

• Main Authors: Shih-Hung Chiu, 邱世宏
• Other Authors: Shing-Tai Pan
• Format: Others
• Language: en_US
• Published: 2003
• Online Access: http://ndltd.ncl.edu.tw/handle/13932498545911055810

碩士 === 樹德科技大學 === 電腦與通訊研究所 === 91 === In this thesis, based on genetic algorithm (GA) and steepest descent method (SDM), we present a new sandwich-like algorithm to identify the nonlinear system by Back-Propagation Network (BPN). The weights and bias of neural networks are trained by the sandwich-like algorithm proposed in this thesis. There are three stages in our new algorithm. The first stage searches, by steepest descent method, for a set of more “nice” initial values for the learning of the weights in neural network. In the second stage, based on the initial values obtained from the first stage, the genetic algorithm is used to make a global search of the weights which optimize the cost function of the output of neural network. In the third stage, for speeding up the convergent rate of the learning algorithm, the steepest decent method in used again to search the finial optimal solution of weights. Some examples of nonlinear system including chaotic system is simulated in this thesis. The simulation results show that the learning performance using the sandwich-like algorithm on Back-propagation network is much better than those using other method.