支持向量机算法在化学化工中的应用

由于计算机技术的发展，机器学习（包括线性和非线性回归，人工神经网络，模式识别算法等）已成为处理化学化工数据，总结经验规律，据以预报未知或控制生产过程的常规手段。但是，传统的机器学习算法都以经典统计数学的渐近理论为依据。该理论的大数定理规定：统计规律只有在已知样本数无限多时才显露出来。但化学化工实际工作中已知样本总是有限的。忽视这一矛盾是造成实际计算中过拟合弊病的重要原因。针对经典统计教学这一弱点，Vapnik学派提出了“统计学习理论”和“支持向量机算法”。新算法既能处理非线性问题，又能抑制传统算法（如人工神经网络等）常遇到的过拟合弊病，本专刊中的论文系列工作表明：支持向量机算法在分析化学的多变量校正，数据处理，商品检验，相图和新化合物的计算机预报，新材料制备的实验设计，环境污染的建模和预报，以及分子设计，药物设计等领域的应用都有良好的效果。在多数情况下所建的数学模型较传统算法的结果有更好的预报正确率，这一新算法将会成为化学，化工领域数据处理广泛应用的新计算工具。

Support vector machine applied to chemistry and chemical technology

Since the progress of computer science and technology, machine learning( including linear or nonlinear regression, artificial neural network and pattern recognition methods) has already become widely used data processing tool in the fields of chemistry and chemical technolo-gy. These traditional data processing methods are based on the asymptotic theory of classical statistical mathematics, which implies that the generalization ability can be achieved as the number of training samples approaching infinity. But the real statistical work in the fields of chem-istry and chemical technology can only use finite number of training samples in practical works. Ignorance of this contradiction may induce overfitting problems in practice. In order to solve this problem, Vapnik and his coworkers have worked out a new theory called " statistical learning theory" and a new method of computation called "support vector machine" . This new computation method can be used dealing with nonlinear problems without significant overfitting problem. Our research works published in this issue indicate that this new data mining method can be used for the multivariate calibration, data processing and counterfeit detection in analytical chemistry, computer prediction of phase dia-grams and new compounds, pollution modeling in environmental protection works, experimental design m new materials exploration works and drug design or molecular design, with good results. The mathematical models built by this new method usually exhibit better prediction ability than that built by ANN or other traditional methods. These results indicate that support vector machine is a new data mining method with great potentialities in many fields of chemistry and chemical technology.