Bayesian network structure learning based on an improved genetic algorithm

For the structure learning of the Bayesian network, the existing genetic algorithm is apt to fall into local optimum and has no way to search for the best solution. Therefore we propose an improved genetic algorithm. First of all, we use the mutual information and the Bayesian information criterion (BIC) function to determine the initial Bayesian edge set and then calculate individuals and form the initial population with chaotic mapping and random processes respectively. Second, we cross multiple columns in the unit of individual column vector and then use a roulette to select an illegal graph and modify it so as to reduce the scope of search space. Finally, we use the Asia Bayesian network and the Cancer Bayesian network to verify the effectiveness of our improved genetic algorithm.