基于蚁群算法的选择性神经网络集成方法

Ant conlony optimization-based approach for selective neural network ensemble

为选择差异度较大、精确度较高的神经网络个体组建神经网络集成,提高神经网络集成的性能,提出一种新的选择性神经网络集成构造方法.该算法采用蚁群优化算法在独立训练的神经网络个体中选择部分组建网络集成,在蚁群优化过程中神经网络个体被选择的概率由信息素和启发因子决定,信息素反映当前神经网络个体的精确度,启发因子反映神经网络个体间的差异度,能有效提高系统的搜索效率和预测精度.实验结果表明,该算法构造的神经网络集成使用了较少的网络个体,而预测误差均好于传统的Bagging和Boosting算法.

A new approach was presented to improve the performance of selective neural network ensemble by choosing the appropriate individuals that are accurate and diverse from candidate neural networks. Ant colony optimization algorithm was employed in which the selective probability depends on the pheromone and heuristic information. The pheromone is re-specified according to the accuracy of individuals while heuristic information indicates the diversity of individuals. The experiments on typical date sets show that this approach yields ensemble with smaller size while achieving much better performance, compared to the traditional Bagging and Boosting algorithm.