【摘要】 采用改进的遗传算法(IGA)和BP人工神经网络相结合的方法,研究了50个酚类化合物的麻醉毒性和分子结构之间的相关性,并与单纯用BP人工神经网络建立的模型进行比较.结果表明,该方法克服了人工神经网络训练中的局部最优问题,采用最优交叉和变异等遗传策略,有效地解决了收敛过程中的振荡问题,所得模型的训练精度和预测精度均优于单纯的BP人工神经网络QSAR模型.更多还原

【Abstract】 Using an improved genetic algorithm combined with a BP algorithm a new method IGA-BP,QSAR model was developed that links molecular structures of a set of 50 alkylated and/or halogenated phenols with their polar narcosis toxicity.Comparing the results of IGA-BP with those of BP algorithm demonstrated that the method of IGA-BP overcame a local optimal solution which often occurs in the training process of artificial neural network.Genetic strategies, such as optimal crossover and optimal mutation,which can efficiently extinguish the swing problem in a convergent process,were employed that resulted in lower errors than those of the BP algorithm.更多还原