Improved deep reinforcement learning for robotics through distribution-based experience retention