Abstract: To the question of drawbacks in learning Bayesian network with genetic algorithm, an immune genetic algorithm was proposed and used to learn the structure of Bayesian network, which combines the constraint based approach with scoresearch based approach. The algorithm can avoid generating illegalstructures; by means of the property of Markov equivalence, the immune operators maps the search space from skeleton space to Markov equivalent class space. The experiment data show that the search space was decreased, compared with those of the genetic algorithm search in direct acyclic graph space, the convergence speed and the efficiency were improved.

Key words: Bayesian network, structural learning, Markov equivalence, immune genetic algorithm, conditional independence test