Contingency-Aware Exploration in Reinforcement Learning