Parameter-Free Reduction of the Estimation Bias in Deep Reinforcement Learning for Continuous Control