Traditional Genetic Algorithms initialize their individuals stochastically and generate too many eccentric and homogeneous individuals. This will cause premature convergence and slow down the speed. In this paper this important problem is studied via the complementary-parent strategy of initializing population in PGA. We analyze it and conclude that the limiting probability of the traditional mutation operator based on this strategy is 1/|HL|2 higher than on the traditional one. This strategy is more efficient by applying to Coarse-grained parallel genetic algorithm which develops the parallelism among populations. For PGA, we also discuss the reproductive capacity of excellent schemata based on the schema theorem and demonstrate the global convergence using homogeneous finite Markov chain. Finally, we present an improved algorithm named GACPS. With some of typical unsymmetrical functions tested, simulation show the quality of GACPS is much higher than PGA on precision, stability and convergence rate.