Research Of A Bayesian Network Structure Learning Algorithm

There is a mount of uncertainty in real world, so uncertain reasoning is formalily with reality. How making use of uncertainty to reason is the main study field.The probability can estimate the uncertainty. From the relative knowledge of the probability we know some probability can be displayed by the total probability, that is to say we can use the known knowledge to deduce the unknown knowledge, this is the essential of bayes reasoning. However with the enlarge of the varient number, the total probability will became bigger, it do not fit computation. While putting up with the bayesian network, it uses the independence of the varience to simplify the process of calculation, this decrease the amount of conditional probability. The Bayesian network is the directly graphic representation, it vividly describe the corresponding field thing's degree of relative. The Bayesian network can reasoning and predict some thing, it is a main method that dealing with uncertain fact. Usually construct a fitting network by expert according to his knowledge.However human construct network is time-consuming and labor-consuming for complicated network. If an expert was careless, he will construct a wrong network. We can acquire a great deal of data by observation in the world, the computer was suitable to large computation, so the people start to consider the new constructing network way ――machine learning. Learning Bayesian network by computer can save human resource, improve the work efficiency, be advantageous to the scope of Bayesian network application. The textual work was completed in carry on the national high technology research development plan of China (863 plan) "Research and Development Of Integrate Circumstance Of Intelligent System". The original Bayesian network learning control in the platform was carried out by EGA (a kind of Bayesian network learning method based on Genetic Algorithm), however we find some weakness existing in EGA during developing process, such as evolving speed is slow, hard to convergence, the result that learned is not very ideal. So we hope to find out a better method. This thesis carries on the thorough research to the Bayesian network and its structure learning, consulting some reference, especially doing the analysis and summaries to some research methods of the current Bayesian network structure learning, I think using GA(Genetic Algorithm) to learn network in a great deal of study method is fitter, because the GA operating on the gene code of the network structure, and the random mechanism can break the trap of the local smallest, obtain the better solution. So we hope making use of the good characteristic of GA, and avoid the disadvantageous factor of solving the problem. According to above-mentioned consideration this thesis use the immunity evolution algorithm to learn the Bayesian network, the immunity evolution algorithm is a kind of algorithm that going into the immunity mechanism, it use "vaccination"to operate on individual. By analysis to the knowledge and experience of some field we can abstract vaccination, it is a method which can avoid forming a pattern unwanted, then use these vaccination to operate on the individual, this was useful to the evolution of the community. The key of the immune evolution algorithm is to abstract fitness vaccination. This thesis aiming at the network abstracts several vaccination, by experiment we find the network that learned by immune algorithm is better than the network learned by EGA and the convergence speed is faster (especially the large sample). However there is still the random mechanism in the algorithm after all, so sometimes it will lead the algorithm stop early. But this is the common problem using random mechanism. Taking the average value or best value through calculation of many times, and adjusting parameters will neglect the influence of bad factor, then obtaining good network.