Research On Classification Algorithms Based On Learning Automata

With the rapid development of information technology in recent years,social information manifests an explosive growth. Data mining plays animportant role in retrieving and utilizing data efficiently. As a basictechnique, data classification provides fundamental methods for theautomatic recognition as well as filing of information and plays a key rolein data mining. With the diversification of information sources andmassive scale of data nowadays, the category information of data and thedata processing steps are more liable to be corrupted by noise. However,the task of realizing robust data classification still needs to be furtherresearched. To improve the performance of data classification in stochasticenvironments, we deeply research on the convergence property,reinforcement scheme and updating rule of continuous-parameterLearning Automata (LA) classifier and propose correspondingclassification methods based on the theory of LA.Firstly, we take a deep research on the Generalized Learning Automata (GLA) classification algorithm. Feasible solutions are given after reasonsthat may cause low learning speed and unstable convergence are analyzed.This dissertation combines the GLA with the heuristic rule of variable stepsize and yields an improved learning algorithm based on adaptive step size.The new algorithm makes use of the relativity among the stochasticgradients and adaptively adjusts the step size during the learning process.Besides, robustness is achieved by filtering unreasonable updates througha proper threshold. Meanwhile, detailed theoretical analysis for thecharacteristics of the step size is shown in this paper. The experimentresults show that better learning speed and stability are simultaneouslyachieved by the modified algorithm. Furthermore, it retains the originalnoise-tolerant ability of GLA, thus providing a better method forclassification in noisy environments.Secondly, the characteristics as well as drawbacks of traditional splitsearching techniques in the nodes of oblique decision tree are analyzed. Totackle the problems of disturbed fitness evaluation which are resulted fromdata set sampling in oblique decision tree inducing algorithms, we proposean oblique decision tree inducing algorithm that uses Continuous ActionLearning Automata (CALA) as the split optimization technique. With asingle-dimensional optimal split as the initial state of CALA, the algorithm shows more stable searching ability and is able to keeprelatively high accuracy in environments with serious evaluation noise.The experiment results show that the method is viable and effectivelyimproves the tolerance of evaluation noise for oblique decision treeinducing algorithm.