In the process of multi-granulation rough set's weight calculation and the condition attributes reduction, the equivalence partition produced by single attribute is usually ignored. Therefore, the attribute reduction problem of multi-granulation rough sets is transformed into the discrete multi-objective optimization problem by introducing the idea of Pareto optimality, in which both resolution of dependability based on equivalence relation and attributes’ significance are taken into consideration. For this optimization problem, a swarm intelligent optimization algorithm with complex network based population structure is designed, in which the non-dominant solution set for individual is introduced to balance the local optimum and global optimum, and the mean-variance based genetic operator is also designed to increase the diversity of the population. The optimization is conducted based on the rough set decision tables which are obtained from the test data sets in UC Irvine Machine Learning Repository, some other multi objective intelligent algorithms are also introduced as comparison, then based on the reduction results, the multi-layer perceptron is introduced to classify the samples in data sets, and the validity of the proposed algorithm is verified. The results show that: 1) The algorithm proposed in this paper shows better performance in multi objective attribute reduction; 2) The algorithm of multi objective rough set attribute reduction combines the advantage of knowledge resolution and knowledge granularity well, and also improves the classification accuracy of the data sets.