An improved genetic algorithm is proposed for solving premature convergence. Firstly, the population is divided into several sub-populations by the minimum spanning tree clustering. Then, the genetic operation is performed among individuals within sub-population which ensures the evolution direction and speed, and that among individuals between different sub-populations which provides diversity by avoiding inbreeding. The experimental results on 23 benchmark functions using binary and real-valued representations show that the proposed algorithm has better convergence and faster speed to get the optimal solution.