论文部分内容阅读
在基于梯度下降原理的ＢＰ网络学习过程中，权值的获取方法是采用某个确定的权值变化规则，然后在训练中逐步调整，最终得到一个较好的权值分布。但它往往会因非线性多极值目标函数而陷于局部最优解。本文采用全局寻优的遗传算法（ＧＡ）和基于梯度下降的局部寻优反传算法（ＢＰ）相结合来训练网络，使网络的连接权在不断迭代过程中自适应演化。通过在ＮＨ地区利用井旁道地震特征参数外推重建井底以下声波曲线的实践，表明这种演化学习方法可以克服传统方法的不足，而且还能避免训练中的“伪学习”现象，提高网络的推广预测能力。
In the BP network learning process based on the gradient descent principle, the method for obtaining the weight is to adopt a certain rule of weight variation and then adjust gradually during the training, and finally obtain a better weight distribution. However, it tends to get trapped in the local optimal solution due to the non-linear multi-polar objective function. In this paper, a global optimization genetic algorithm (GA) and a gradient descent based local optimization backpropagation algorithm (BP) are used to train the network so that the connection right of the network evolves adaptively during the iterative process. The practice of extrapolating and reconstructing the acoustic curve below the bottom of the shaft by using the seismic characteristic parameters of the by-pass in the NH area shows that this evolutionary learning method can overcome the deficiencies of the traditional methods and can also avoid the phenomenon of “pseudo-learning” in training and improve the network The promotion of forecasting ability.