Accelerating Reinforcement Learning for Robot Controls Using Interim Rewards and Master/Slave Computing

Reinforcement-Learning (RL) has become a bigger task in engineering. The advantage of RL is the ability to learn without known datasets, only with rules included in an environment. If the task becomes bigger, the learning phase grows enormously. Two approaches for accelerating the time of learning are discussed in this work. Introducing interim rewards offers the possibility to split a big task into several smaller tasks. With the master/slave approach, multiple instances learn individually and the knowledge of each of them is merged afterwards into one behavior strategy. Both of these approaches are demonstrated by means of a case study.