Robust learning for autonomous agents in stochastic environments