【摘要】 设计了一种新颖的、表达能力很强的多S型函数,以其作为单输出神经元的活化函数所构建的多S型前馈网(MS-FN)分类器,结构简洁,训练速率高,由此建立的多分类模型有很强的判别能力,且具有优良的泛化性和稳定性.还推导了相应的Levenberg-Marquart(LM)训练算法,并应用于胺类有机物急性毒性的定量构效关系,所建模型样本预报正确率高.实例表明MS-FN分类器是化学模式识别的一种有效工具.更多还原

【Abstract】 A novel multi-sigmoid function and the correspondi ng multi-sigmoid feed-forward network(MS-FN)classifier taking the multi-sigm o id as single output unit activation function were proposed in this paper. Then L evenberg-Marquart(LM)algorithm with high training speed ada ptable to the network structure was deduced . This compact structure was used in multiple classification . Fine multi-classification performance,excellent extensive and steady capability were found from its application to the Quantitative Structure-Activity Relationship(QSAR)of amines acute toxicity by using different activation functions in the network ou tput layer. This example illuminated that the MS-FN classifier was an effective tool for chemistry pattern recognition.更多还原