Based on improved evolution strategies with single-gene mutation, elitist reproduction, and descending strategy parameters, a cooperating multi-population evolution strategy is proposed for the optimization of multi-modal function. The subpopulation's probability of evolution, stopping condition, and criterion of convergence to local optima are presented. A new method named as valley searching to distinguish between same-peak optima and different-peak ones is proposed, and determining the radius of niche or peak is avoided which has been a hard problem for fitness sharing genetic algorithms. Simulation results on a set of benchmark functions show that the algorithm can properly find all the optima.