Abstract: Aiming at the problems that there are many cognitive users in the dynamic spectrum assignment issue in cognitive radio,and the traditional optimization algorithms cost amounts of convergence time. A multi-level dynamic spectrum allocation algorithm is presented based on branch and bound theory.This paper first build a model of DSA system with multi-level based on user demand, and then the problem was formulated as a functional optimizing by means of the graph-coloring theory, finally The branch and bound optimization algorithm is introduced to the model, the algorithm put all the practicable solution space constantly dividing into smaller and smaller subsets, by this way the dynamic spectrum allocation is implemented to the model of DSA system with multi-level.The simulation experiment is taken with comparison on the basic genetic algorithms,through the analysis of the amount of the cognitive user access and the system profits, the simulation results show the superiority of the classification model in dealing with the problem of DSA with multi-level. And the algorithm proposed has an relatively small computation complexity, which is of a relatively high application value.