基于学习自动机的容噪模式分类  

Noise-Tolerant Pattern Classification Based on Learning Automata

作　　者：刘晓[1] Liu Xiao(AVIC Xi'an Aeronautics Computing Technique Research Institute,Xi'an 710065,China)

机构地区：[1]航空工业西安航空计算技术研究所,陕西西安710065

出　　处：《航空科学技术》2020年第10期75-80,共6页Aeronautical Science & Technology

摘　　要：在模式分类学习问题中,训练数据中的标注差错(也称类别噪声)对分类器的性能有很大的影响。本文将一种新近提出的连续动作学习自动机(即聚焦区间学习自动机)应用于针对类别噪声的容噪学习问题。分类器采用简单的单隐层前馈神经网络,利用一个由这种学习自动机组成的自动机团队,对神经网络的权值参数进行学习。通过广义异或问题和Iris数据集的仿真试验,将该算法与两种基于群体搜索的优化算法——粒子群优化(PSO)和差分进化(DE)进行了比较研究。结果表明,新算法具有更好的容噪学习性能。In the learning problem of pattern classification, the label error(also called class noise) in the training data can severely impact the performance of the classifiers. A recently proposed continuous-action learning automaton, i.e.,the focused interval learning automaton is applied to the noise-tolerant learning for the class noise. The classifiers adopt simple single hidden-layer feed-forward neural networks. A team of such learning automata is used to learn the weight parameters of the network. Simulations are carried out which employ the new algorithm and two populationbased optimization algorithms, particle swarm optimization(PSO) and differential evolution(DE), respectively, on the generalized XOR problem and the Iris dataset. The simulation results indicate that the new algorithm can obtain better noise-tolerant learning performance compared with the PSO and DE.

关 键 词：模式分类 类别噪声 容噪学习 学习自动机 连续动作学习自动机 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]