Noise-Tolerant Pattern Classification Based on Learning Automata

Release date:2020-11-06

Liu Xiao*
Xi'an Aeronautics Computing Technique Research InstituteXi'an 710065China

Abstract: 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.

Key Words: pattern classification; class noise; noise-tolerant learning; learning automata; continuous-action learning automata

©2019 Chinese Aeronautical Establishment (CAE) All rights reserved. ICP:14045040