混沌优化神经网络方法及其在地震多属性研究中的应用

A method of chaos optimization neural network and its application in multi-attribute seismic research

针对传统的基于梯度下降法BP神经网络中存在非线性多极值目标函数易陷于局部最优解的问题,提出了一种权值学习混沌优化的神经网络方法。非线性动力系统具有初值敏感性、遍历性的特性,采用基于混沌和梯度反传训练相结合来训练网络,可以使网络的连接权在不断迭代过程中自适应演化。实际过井地震剖面地震多属性研究实践表明,所提出的混沌优化学习方法可以克服传统方法的不足,提高预测能力。

In the traditional neural networks based on gradient decline, nonlinear object functions with multiple extrema are easy to get stuck in the problem of local optimization solution. Therefore, a new method of chaos optimization neural network with weight learning is introduced in this paper. Because of sensitive dependence on initial conditions and ergodicity of nonlinear dynamic systems, a combination of chaos with gradient backpropaga- tion can be used to train the neural network, by which the weight of a network may be in self-adaptive evolution during continuous iteration. A case of multi-attribute research on seismic sections across wells shows that this learning method of chaos optimization can avoid shortcomings of traditional approaches and improve prediction ability of a neural network.