Accelerating Reinforcement Learning by Composing Solutions of Automatically Identified Subtasks