Structure Learning of Bayesian Networks Using Heuristic Methods
Alireza Sadeghi Hesar, Tabatabaee Hamid, Mehrdad Jalali, Ghoochan Branch · 2012
In this paper, we introduce bayesian artificial networks as a causal modeling tool And analyse bayesian learning algorithms. Two important methods of learning bayesian are parameter learning and structure learning. Because of its impact on inference and forecasting results, Learning algorithm selection process in bayesian network is very important. As a first step, key learning algorithms, like Naive Bayes Classifier, Hill Climbing, K2, Greedy Thick Thinning are implemented and Are compared based on accuracy and structured network time. Finally, the best of learning algorithm will be proposed. We work with a database of observations (monthly rainfall) measured for the years 1985-2010 in a network of 22 stations in the (Rzavi, Shomali And Jonoubi) Khorasan provinces and with the corresponding gridded atmospheric patterns generated by a numerical circulation model. A Bayesian network or BN is a model that reflects the states of real world and describes how those states are related together by probabilities. Using this framework, the inherent structure of different processes can be interpreted. Although Bayesian Networks provide a means for inference but finding the structure of the networks remains an NP-hard problem. The reason for this is that there is an enormously large number of ways in which the network nodes can be linked to each other. In order to mitigate this problem, a number of algorithms have been proposed. Like, the Naive Bayes Classifier, K2, Local K2, Greedy Thick Thinning or GTT algorithms and etc. The main purpose of this paper to determine the algorithm which produces the Bayesian network with the highest predictive accuracy, and is constructed in the least amount of time. The first part of the article provides a brief introduction of bayesian networks. Then, experimental data sets and methods of normalization is reviewed. In the next section, the main bayesian learning algorithms is introduced and inference methods of them is descripted. In the last section, Results related to the implementations and comparison of algorithms is expressed (1)(2).