基于旋转曲面变换PSO算法的神经网络用于胺类有机物毒性分类

神经网络模型能有效地模拟非线性的输入输出关系。本研究应用三层前馈网络对51种胺类有机物进行了结构-毒性关系的分类研究。常规的神经网络权值训练算法,例如误差反传算法,存在着收敛速度慢,容易陷入局部极值点等问题。因此提出旋转曲面变换粒子群优化算法,将被优化函数的局部极小点变换为全局最大点,同时不改变比局部极小点的值更小的区域的函数形状。此方法和粒子群优化相结合,能使待优化函数跳出局部极值点,提高训练神经网络权值的效率。实验结果显示,基于旋转曲面变换粒子群优化算法的神经网络,权值训练过程收敛速度较快,且自检误差和预报误差都较小,是一种有效的胺类有机物毒性分类方法。

Classification of Toxicity of Amines Using Neural Networks Based on Rotate Surface Transformed Particle Swarm Optimization Algorithm

Neural networks (NNs) have become one of ideal tools in modeling nonlinear relationship between inputs and desired outputs. In this study, 3 layers of feed-forward network have been used for structure- toxicity relationship analysis of 51 kinds of amines . However, the training of NNs by conventional back-propagation (BP), i.e. the BP-NNs, has intrinsic vulnerable weakness in slow convergence and local minina. In this work, rotate surface transformed particle swarm optimization(PSO) algorithm was proposed to train NNs, which was named HPSO-NNs. Rotate surface transformation (RST) method was proposed to overcome the defects . RST method transforms local minima to global maximum and keeps the values of the function to be optimized unchanged where the value is lower than local minima, then RST can help the networks jump out from local minima and improve PSO train efficiency. Experimental results illuminat that HPSO-NNs accuracy of classification is improved, and error of prediction is decreased, which shows that HPSO-NNs is a good method for the classification of toxicity of amines.