Research of AdaBoost robustness based on Learning Automata

摘要

AdaBoost is one of the widely used techniques in the area of pattern recognition. It searches the samples that are classified incorrectly and subsequently pays more attention to them at the next iteration. While the dataset is noisy, according to the iterative mechanism of AdaBoost, the noisy samples may be paid more attention at, which might lead to bad behavior. As a method of reinforcement learning, learning automata could search the optimal state adaptively in a random environment. In this way, we believe the algorithm based on learning automata shows good behavior towards noisy dataset. In this paper, we improve the performance of AdaBoost based on learning automata by adjusting the weight through a stochastic optimization method. The experiments indicate that the algorithm is able to restrain the negative effect caused by noise effectively.