Among various frameworks of intelligence, in general, feed-forward perceptron neural networks (FPNN) is a useful and common method, because of the network\'s ability to approximate highly nonlinear functions. Similarly, among various paradigms of learning, evolutionary-based algorithms such as genetic algorithms (GA) have gained increasing interest in recent years due to their ability to locate globally optimal solutions in nonlinear, noisy and uncertain problem domains. Here, we propose a cooperative co-evolutionary strategy for finding weights and structure of FPNN simultaneously. The new algorithm allows for separate populations of weights and structures of neural networks to coexist and cooperatively evolve thru two separate genetic algorithms. The proposed algorithm is simulated in RoboSoccer multi-agent environment, and is used for learning the \"ball interception\" skill of robot soccer players. Also, the convergence properties of the new algorithm are statistically compared with two other approaches as well as standard back propagation (BP) algorithm. Simulation results indicate that the proposed co-evolutionary approach is superior in terms of consistently finding improved solutions.

Keywords