一种基于神经网络的仿真优化方法

Optimization via Simulation Based on Neural Network

为提高仿真优化问题求解效率,提出了一种基于神经网络的仿真优化方法。利用神经网络对非线性输入输出关系的逼近能力,由神经网络输出值代替仿真结果以减少所需仿真次数。按照提出的3种样本选择策略,由仿真模型产生一定数量的样本,借助广义回归神经网络在学习速度、网络稳定性、参数选取方面的独特优势,对样本进行训练,生成能够反映仿真模型输入输出关系的回归曲面,以实现用神经网络输出值代替仿真结果,利用优化算法对回归曲面进行寻优。通过对典型测试函数进行实验,证明了方法的可行性和有效性。

To improve the efficiency of optimization via simulation （OvS）, an OvS method based on neural network is proposed. Taking advantage of the approximation ability of neural network to nonlinear input-output relationship, neural network＇s outputs are used as substitutes for simulation results to reduce the required simulation runs. Samples are generated by simulation according to the three proposed samples selection methods. Owning to its advantages on learning speed, network stability and parameters selection, generalized regression neural network （GRNN） is adopted to train the samples. The trained GRNN forms a regression surface that represents the relationship between simulation inputs and outputs, which makes it feasible to use GRNN output as substitutes for simulation runs. Optimization algorithms are applied to search for the best solution on the regression surface. Experiments are carried out with some typical test functions, and the feasibility and effectiveness of our method are demonstrated.