Metrics and continuity in reinforcement learning