Based on unconstrained optimization and genetic algorithm, this paper presents a constrained genetic
algorithm(CGA) for learning Bayesian network structure. Firstly, an undirected graph is obtained by solving an unconstrained
optimization problem. Then based on the undirected graph, the initial population is generated, and selection, crossover and
mutation operators are used to learn Bayesian network structure. Since the space of generating the initial population is
constituted by some candidate edges of the optimal Bayesian network, the initial population has good property. Compared
with the methods which use genetic algorithm(GA) to learn Bayesian network structure directly, the proposed method is
more efficiency.