Risk-averse reinforcement learning