This study derived an expectation-maximization (EM) algorithm for estimating the parameters of multidimensional item response models. A genetic algorithm (GA) was developed to be used in the maximization step in each EM cycle. The focus of the EM-GA algorithm developed in this paper was on multidimensional items with "mixed structure." Simulated item response data were generated and then estimated by a computer program based on the EM-GA algorithm. The simulation results demonstrate that the EM-GA algorithm is a very promising approach in estimating multidimensional item response model parameters.