基于冲突指导的神经网络预测方法

研究基于冲突指导的神经网络预测技术. 利用离散约束最优的拉格朗日乘数理论， 通过采用前馈方法得到近似梯度的模拟退火技术， 避免了盲目接受试验点. 利用松紧策略加快了搜索的收敛速度. 实验结果表明， 训练误差和预测误差都有很大改善.

Neural networks prediction method based on violation guide

The neural network prediction technique based on the impulse guidance is discussed. On the basis of the theory of discrete constrained optimal Lagrange multiplicator, we avoided to accept test points sightlessly with the simulated anneal technique from the approximate grades through the forward propagation method, and accelerated the search speed using Relaxed & Tighten strategy. The experiment result shows that the training error and the prediction error have been improved greatly.