动态神经网络分类器主动学习算法及其智能控制应用

APPLICATION OF ACTIVE LEARNING ALGORITHM FOR DYNAMIC NEURAL NETWORK CLASSIFIER AND ITS INTELLIGENT CONTROL

针对动态神经网络分类器训练时采样时间长、计算量大的问题,提出一种动态神经网络分类器的主动学习算法。根据主动学习AL(Active Learning)算法中一种改进型不确定性采样策略,综合考虑样本的后验概率及其与已标记样本间的相似性,标注综合评价得分值较小的样本,将其用于对网络分类器的训练。通过Sobol’敏感度分析法,神经网络适时地增加敏感度值较大或删减敏感度值较小的隐层神经元,以提高其学习速率,减小输出误差。分类器训练仿真实验结果表明,与被动学习算法相比,该算法能够大大缩短网络分类器训练时间,降低其输出误差。将该算法用于液压AGC系统中,实验结果表明,该算法可实现系统中PID控制器参数的在线调节,提高了厚度控制精度,以此验证了该算法的适用性。

In view of the long sampling time and large computation amount in training process of dynamic neural network classifier,we put forward an active learning algorithm for dynamic neural network classifier. According to an improved uncertainty sampling strategy in active learning algorithm,and considering both the posterior probability of the sample and the similarity of the marked samples,we annotated the samples with smaller comprehensive evaluation score and applied them in network classifier training. By sobol'sensitivity analysis method,the neural network timely increased or pruned the hidden layer neurons with larger sensitivity value or smaller larger sensitivity value in order to improve the learning rate and reduce the output error. Results of simulation experiment of classifier training showed that compared with the passive learning algorithm,the proposed algorithm could greatly shorten the network classifier training time and reduce the output error.Applying the algorithm to hydraulic AGC system,the experimental results showed that it could realise the online adjustment of PID controller parameters in system,and improve the precision of the thickness control,these validated the applicability of the proposed algorithm.