Study Of The Bayesian Network Structures Incremental Learning

In the 1980s, at the aspect of solving the uncertain problems on artificial intelligence, a new subject, named Bayesian Network, which contains the probability theory, statistical theory and graph theory comes into being. Due to using the clearly and vividly graph to reveal the statistical relationship among the variables in the researching objects, Bayesian Network can make the expression of complicated probability distribution in a clearly and compactly method. So, for the prominent advantages, Bayesian Network becomes effective tool in uncertain inference and data analysis.The traditional Bayesian Network structure learning algorithms are batch learning algorithms, which execution time is too long and which need huge storage space when in the condition of dynamic training data set. To improving the efficiency of the learning algorithms and making up for the disadvantages of batch learning algorithms, the main work of this paper and the innovations of this paper are as follows:Base on the particle swarm optimization and genetic algorithm, a new searching algorithm named mixed PSO search (MPS) algorithm adapt for Bayesian Network structure learning is proposed. And then using the experimental data to prove the MPS algorithm is valid.Then, I research the incremental method for learning, and there are two optimal functions in the incremental method for learning, which are WTUD function (When To UpDate function) and SSS function (Shrink Search Space function). The WTUD function should determine whether the structure should be updated and when the structure should be updated. The SSS function should determine the search space of the algorithm. Based on this incremental method, a new incremental MPS algorithm which induces the incremental method to MPS algorithm is proposed. And then the experimental data proves that the MPS algorithm is valid.The traditional Bayesian Network structure learning CL algorithm and B algorithm are batch learning algorithms. Based on the traditional CL algorithm and B algorithm, two new algorithms named iCL and iB, which induce the incremental method, were be proposed. The via making the experiment, we find that these two new algorithm, iCL and iB, achieved the demand expected, that is reduce the algorithm executing time and save the memory space needed.