Block Climbing Learning Algorithm For Tractable Bayesian Networks Model Based On Pruning

The core of AI lies in how to express the existing things and analyze or reason the external things in the understanding of the existing things.The Bayesian network(BN)model can well reflect the correlation and independence between the attributes of things.It is the combination of graph theory and probability theory.It emphasizes on use the direct acyclic graph and conditional probability to describe the dependence between variables,and on this basis,it carries out the probability reasoning between variables.Due to its unique expression form for uncertain knowledge,rich probability expression ability and learning of comprehensive prior knowledge,BN has become one of the most effective models in the field of uncertain knowledge representation and inference.This paper focuses on the problems of high space complexity of searching for the best candidate BN,easy to fall into local optimization of learning results,and complexity of controlling BN inference in the climbing learning algorithm of BN model structure,and proposes a block climbing learning algorithm of tractable BN model based on pruning.Research contents of this paper include:Firstly,to avoid the problems of high complexity of candidate BN search space and local optimization when climbing step by step in traditional climbing method,a PBBN(based-pruning learning for Bayesian networks with block)algorithm is proposed.First,in order to reduce the complexity of search space of candidate BN,a based-pruning search strategy is proposed to obtain the possible combination of parent nodes of each node in the network.Then,based on the pruning results,a block climbing strategy of candidate BN model is proposed,including the selection of block size and the making of block search strategy.Finally,PBBN algorithm is designed and implemented.On the Sachs,child and alarm standard networks,several evaluation indexes are used to verify the good performance of the proposed PBBN algorithm by comparing with other BN model structure learning algorithms.The algorithm not only reduces the search space of candidate networks,but also overcomes the difficulty of learning from high probability candidate networks when climbing step by step.To some extent,it avoids the problem of easily falling into local optimization.Secondly,to solve the high complexity problem of BN model reasoning,on the basis of PBBN algorithm,a block learning TPBBN(Tractable PBBN)algorithm for BN model is proposed.Specifically,combined with the elimination tree learning idea of tractable BN model step by step,this paper studies how to learn effective elimination tree in each block climbing and make the learned candidate BN model tractable.TPBBN learning algorithm is proposed by making corresponding learning strategy of BN structure with block.Then on several standard BN models,the performance of TPBBN algorithm is verified by comparing with other learning algorithms of BN model.The algorithm not only guarantees the excellent network structure,but also satisfies the tractability of reasoning based on the BN model.Thirdly,Bayesian network provides a model of human brain reasoning process.According to the strong reasoning ability of Bayesian network,the learning algorithm proposed in this paper is applied to the reasoning learning of arithmetic concepts with which people are familiar in everyday life.First,the Bayesian network model of arithmetic concept is given,then the learning method of tractable Bayesian network of arithmetic concept is proposed,Next the giving strategy of evidence concept is analyzed emphatically,and then the concept reasoning method based on this model is proposed.Finally,by comparing and analyzing the cognition of human beings in arithmetic concepts,the experimental results verify the rationality of the algorithm that has been applied to the learning and reasoning of arithmetic concepts.