In order to overcome the shortcomings of some evolutionary algorithms with slow or premature convergence, such as standard genetic algorithm (SGA), a novel dual population genetic algorithm with learning scheme (SDGA) is proposed in this paper. Based on the SGA, the population is divided into two groups. The offspring produced by crossover can join in one of the two groups, or be discarded by fitness comparisons after learning or mutation. With learning scheme and suitable member updating rules, the proposed algorithm is capable of outstanding global optimization. Numerical results show that the SDGA has a high success rate and low computation consuming for global optimizations. Moreover, the SDGA can give high accuracy solutions for high dimensional problems with lower computing costs.