The Research On Optimum Methods Of Monitoring Incident Detection Using Bayesian Model

Bayesian network (BN) is a probability-based uncertainty reasoning network. It can be represented in the form of a directed acyclic graph (DAG). We can use it to learn and reasoning in the limited conditions of uncertain information. It has many important applications in artificial intelligence in dealing with the uncertainty of information, such as intelligent computer science, industrial control, medical diagnostics and other fields.In recent years, Bayesian network has been more and more applied in modeling monitoring events in the surveillance area, as a result, more and more optimization methods to learn Bayesian network models have been put forward, it has already been a research hotspot. Among the methods of BN optimization, the most commonly used parameter estimation algorithms include maximum likelihood estimation, Bayesian estimation and Maximal Margin and other methods. Expectation Maximization (EM) algorithm is the most common one in dealing with incomplete information. EM algorithm is relatively simple and stable, and it has a good effect on learning Bayesian networks. But in the case that the search space is too large, it is always convergent to local convergence point. And the EM algorithm is aimed at adjusting the target model and the ideological goal of the algorithm parameters, the model has a low ability in adjusting its structure, so it can only play a limited role in the optimization in lack of expert knowledge in the large Scale model. Therefore, considering the high dimension of the parameters and the necessary of adjusting the structure of the Bayesian network model, we take advantage of the genetic algorithm which is more suitable to adjust a wide range of parameters.Meanwhile, as the search space of the Bayesian network is too large, the learning algorithm is easy to achieve local convergence and some other issues, this paper combined the prior knowledge which is obtained from the structure-parameter restrictions with the genetic algorithm to optimize the Bayesian network. It can ensure the accurate learning from a real set of training data. Structure-parameter restrictions in Bayesian network can not only guide the initialization of the structure and parameters of the network, but also can be used in the optimization process to constraint the parameters and the structure to ensure the right direction of the learning algorithm. Experimental results show that the algorithm can jump out of local optimum and global optimum. And it can increase the value of evaluation function more than 6% while ensuring the same convergence rate.We also considered combining genetic tabu list with genetic algorithm to further optimize the learning algorithm, in which the searched areas are noted to avoid repetitive search, as a result, it can effectively improve the speed of convergence.