A LEARNING AUTOMATA-BASED TECHNIQUE FOR TRAINING BAYESIAN NETWORKS NABI ALLAH REZVANI Soft Computing Laboratory Computer Engineering and Information Technology Department Amirkabir University of Technology Tehran Iran nrezvani@aut.ac.ir MOHAMMAD REZA MEYBODI Soft Computing Laboratory Computer Engineering and Information Technology Department Amirkabir University of Technology Tehran Iran mmeybodi@aut.ac.ir ABSTRACT One of the most important challenges of Bayesian networks is training an optimal network based on existing training samples. We propose two Learning Automata-based methods for training parameters and structure of the network. Parameter training method is an incremental method which performs training and testing simultaneously and has lower computational cost than enumerative or search based parameter training methods. The structure training method uses a guided search scheme and avoids getting stuck in local maxima. This outputs a network that improves classification accuracy. We could also use both these methods together to train the network. Results indicated that this combinational method further improved classification accuracy while still kept computational cost rational. KEY WORDS Bayesian Networks, Training, Learning Automata 1 INTRODUCTION Bayesian networks are extensively used as classifiers [1][8][15][16][17]. One of the most important challenges of these networks is training an optimal network based on existing training samples. Training problem consists of two sub-problems: training parameters and training structure of the network. Different researches have been done to improve training methods of Bayesian networks. On parameter training, methods of parameter training have been revised to approach the real goal of classification. Maximum Conditional Likelihood estimation is an example such methods [15]. This sort of estimation involves a search process to find optimal estimations for the parameters. Research works have also been done to improve structure training methods. In these researches network evaluation function is modified and search methods have been proposed and redefined to find optimal network based on the evaluation function [16][17][18]. These methods perform a local search with algorithms like gradient descent or genetic search. These search schemes are blind and heuristic methods and are likely to get stuck in local maxima. Using discriminative methods for both parameter and structure training involves heavy and impractical computational cost, since both methods should run a search process and therefore total cost of training will be the multiplication of these two search processes. This problem is also addressed in this paper.