On-line evolutionary computation for reinforcement learning in stochastic domains