ESTIMATION OF CATALYTIC PERFORMANCE BY USING NEURAL NETWORKS

本文从信息科学的观点出发，对提高神经网络在催化性能预测中的可靠性进行了实验研究。针对二元非线性函数的特殊情况，研究了误差反向传播神经网络在学习过程中所遇到的函数值超调问题。实验结果表明，为了使学习尽可能完整，在这些值和输出层单元的输出之间应该保留与计算函数值大小一样大的余量。同时得出结论，训练数据应用于输入层的方式的差异对这些神经网络的学习性能并不重要。

In this investigation, the methods of promoting estimation reliability of neural network for the case of catalytic performance are experimentally discussed from the information science viewpoints. The problem of function value overshooting encountered in the process of learning with error back-propagation neural networks is computationally investigated for the specific case of two-variable nonlinear functions. The experimental results suggests that the allowance of as large as magnitude of calculated function values should be reserved between those values and the outputs from output-layer units in order to make that learning as complete as possible. It is also concluded that the difference in application manners of training data to input layer is not critical for learning performance of those neural networks.