The selection strategy is an import step of the fireworks algorithm(FWA), which directly affects the convergence efficiency, the convergence accuracy, the sensitivity to the initial value and the ability to jump out of the local optimal. Therefore, an improved selection strategy of the firework algorithm(ISSFWA) is proposed, which establishes the concept of spark peak and exploration spark. The improved selection strategy is proposed which selects peak sparks and selects the exploration spark as the next generation of fireworks. The peak sparks take into account the fitness values and relative position of sparks, which ensures that the global optimal spark and the local optimal spark in the neighborhood of the peak spark are selected. At the same time, it avoids duplication of sparks with similar search ability and keeps the firework with strong global exploration ability. And the exploration spark based on the largest distance enhances the ability of global exploration. In the 10 repetition test of standard and increased position deviation test function, the ISSFWA is superior to the PSO, GA, FWA in terms of the best fitness, and superior to the PSO, FWA in terms of average fitness, but slightly inferior to the GA. This result shows that the ISSFWA can enhance the ability of finding the optimal solution, reduce the sensitivity to the initial value, and improve the search efficiency.