本文详细的讨论了在石油勘探方面应用最多的前馈型神经网络。针对前馈型神经网络所用的传统误差反传算法收敛速度慢的缺点，讨论了动量算法、变步长算法、共轭梯度算法和变尺度算法，其中共轭梯度算法和变尺度算法都有较快的收敛速度。文章讨论了神经网络的预测能力，指出多层前馈型神经网络的预测实质上是函数逼近，并从理论上研究了神经网络的光滑性，指出当网络神经元具有连续传输函数时，神经网络相当于一个多元函数，具有连续性。神经网络的这个性质在实际预测中表现为神经网络具有较好的内插能力及抗噪音能力，并用网络逼近函数验证了这一结论。文章还从理论上研究了网络输出与网络权值的关系，指出通过减小网络权的模能适当控制网络输出的波动幅度，从而为通过控制网络权值来提高网络推广性的正则化方法建立了理论基础。 通常人们所用的正则化方法是给目标函数加上一个带有正则化参数的函数。在应用中通过人工方法调节这个正则化参数。为了寻找恰当的正则化参数往往需要反复的试验，这给实验和实际应用带来很大的不便。本文通过贝叶斯方法近似训练样本的真实分布，自动调节正则化参数。文章通过用神经网络逼近函数和用神经网络根据测井曲线预测储层岩性参数，孔隙度和渗透率，对传统方法和正则化化方法作了比较，证实了正则化方法能有效地减小网络权值，提高网络的推广性能。 然而，正则化方法并不能完全解决网络的推广性问题。文章针对使用神经网络解决石油勘探问题的实际情况，提出在石油勘探中使用神经网络预测时出现的误区，并提出相应的改进措施。

This Paper bat around multilayer feed-forward network which is the most widely applied in petroleum exploration and production field. For getting over traditional back- propagation algorithm shortages which is used in feed-forward network, this paper introduce momentum back-propagation method, variable learning rate back-propagation method, conjugate gradient back-propagation method and variable metric back- propagation method, and conjugate gradient method and variable metric method have faster convergence speed. This paper discussed predicting ability to neural networks, and pointed that predicting essential of multilayer feed-forward networks is function approximation, and studied smoothness of neural networks on theory, and come to the conclusion that when neurons of net have consecutive transfer function, neural networks correspond to multi-variables function and are continuum. In real prediction, this character of neural networks present that neural networks have well interpolation ability and anti-noise ability, and validated the conclusion by neural networks approximating to functions. This paper, basing on theory, further studied the relation between output and weight of networks, and come to the conclusion that decreasing model of weight vector can control fluctuate extent of neural networks output, which lay the foundation for regularized multilayer feed-forward network method which improve neural networks generalization ability by controlling weight of network. The regularized method generally used is that object function is added a regularized function which has a regularized parameter. In practice the parameter is adjusted by artificial method. It is necessary for trial and error to find a right regularized parameter, which is a difficult work in experiment and application. This paper presents Bayesian method approximation real distribution of training samples, and auto-adjust regularized parameter. This paper compare variable metric backpropagation method with regularized method through approximating function and predicting reservoir parameter, such as porosity and permeability, from well log, and demonstrated regularized method can effectively decrease weight value of neural networks and approve generalized ability to neural networks. However regularized method cannot perfectly solve generalized problem of neural networks. The paper presents misemploy of neural networks in petroleum exploration and production field and bring forward correspondence improved measures.