关键词: 神经网络, 仿真优化, 局部极小, 样本增补

Abstract: To reduce the time required on simulation runs of traditional optimization via simulation (OvS) method, a design of algorithm is proposed for generalized regression neural network (GRNN) based OvS. Firstly, some inputoutput samples are obtained by simulation, based on which the GRNN is trained and the initial regression surface that represents the inputoutput relationship of the simulation is obtained by the trained GRNN. Then, by adopting pattern search algorithm on the initial regression surface, all possible local minima points are found, including some fake points that are not local minima. Therefore, a method is proposed to identify those fake points so that only local minima points are located. Next, samples replenishing strategies are given to replenish some samples near each local minima, and the GRNN is trained upon the new samples made up of the initial and replenished samples, which can form an amendatory regression surface. Repeat the process of replenishing new samples until the optimal solution of the OvS can be obtained. An experiment is taken on a typical test function, and the results show that our method can solve the OvS problem while effectively reducing the required simulation samples.

Key words: neural network, optimization via simulation (OvS), local minima, samples replenishment