Evaluation of Optimization Methods for Neural Network
Yuto Yasuoka, Yuki Shinomiya, Yukinobu Hoshino · 2016
In this research, optimization methods are evaluated for Neural Network (NN) learning. NN is used as a classifier in image recognition and learning is important integrant to improve performance. Back Propagation (BP) is a typical NN learning method. However, BP depends on the number of NN layers and the default value of NN weights. Therefore, Particle Swarm Optimization (PSO) is expected as a multipoint search algorithm rather than BP. Nevertheless, PSO has some risks to stop in local minimum in a complex problem. Hence, Random PSO (RPSO), in which PSO is modified, is suggested by the researchers. RPSO does not fall into local minimum by velocity control. This paper shows that NN using RPSO is able to achieve a high discrimination rate and a high learning performance can be expected in solving complex problems. Additionally, RPSO does not depend on the number of NN layers.