Bridging the Gap Between Value and Policy Based Reinforcement Learning