基于神经网络的时变参数系统仿真优化方法

时变参数系统的仿真优化问题是一个新兴的研究课题，相比传统仿真优化，时变参数系统对实时性的要求高，而对解的精度要求不高。本文提出将该问题转换为一类神经网络预测问题，并从理论上证明了该方法的可行性。首先，线下构建神经网络模型描述输入参数到最优解的映射关系；然后，利用训练好的神经网络模型线上实时预测最优解。考虑到边界样本对最优解拟合曲面的影响，提出构建中心样本和边界样本，分别训练两个神经网络模型。仿真和实例表明，该方法能够随时变参数的变化实时给出满意解，从而为求解时变参数仿真优化问题提供一种新的解决思路。

Approach to simulation optimization of time-varying parameters system based on neural network

Simulation optimization (SO) of time-varying parameters (TVP) systems is an emerging research topic, which is different from traditional SO problems in that the real-time requirements are very high while the accuracy requirements are relatively lower. This paper proposes to transform the SO problem with TVP into a neural network (NN) prediction problem, and theoretically proves the feasibility of the method. Firstly, training samples are generated by offline SO, and an appropriate NN model is designed and trained to describe the relationship between input TVP and the corresponding optimal solutions. Then, the well trained NN model is used to realize online real-time prediction of the optimal solution. Considering the impact of boundary samples on the optimal solution fitting curve, we propose to train internal training samples and boundary training samples respectively so that two different NN models will be built to better reflect the fitting curve. The simulation and empirical study show that, our approach can provide online real-time satisfactory solutions as TVP changes, which presents an approach to SO problems with TVP.