摘要
将先验知识与神经元网络相结合，可以提高模型的拟合精度和预测能力。本文将针对三层前馈网与单调性先验知识相结合的问题，分析Joerding的惩罚函数法，提出两种新方法：插值点法和有约束优化方法，并成功地应用于原油实沸点蒸馏曲线的仿真，使网络模型在整体和局部上都更贴近于实际对象。
FFN has been widely applied in modeling chemical processes because of its universal approximability. The inclusion of prior knowledge is a means of improving the fit precision and the prediction ability of the modal when trained on sparse and noisy data. As to the three-layer feedforward networks and the prior knowledge of monotonicity constraint, the Joerding's penalty function method is analyzed first. Then two novel methods: interpolation method and constrained optimization method, are proposed. These methods have been applied to modeling the true boiling point curve of the crude oil successfully. The simulation experimental results show that the network models trained by those methods are more close to the actual object in local and whole.
出处
《高校化学工程学报》
EI
CAS
CSCD
北大核心
2001年第4期351-356,共6页
Journal of Chemical Engineering of Chinese Universities
基金
国家自然科学基金资助项目(编号:20076041).
