Self Learning in Flexible Manufacturing Units: A Reinforcement Learning Approach

This paper presents a novel approach for self-learning as well as plug-and-play control of highly flexible, modular manufacturing units. The approach is inspired by recent encouraging results of reinforcement learning (RL) in computer science. However, instead of learning the entire control behavior which results in long training times and requires huge data sets, we restrict the learning process to the supervisor level by defining appropriate parameter from the basic control level (BCL) to be learned by learning agents. To this end, we define a set of interface parameter to the BCL programmed by IEC61131 compatible code, which will be used for learning. Typical control parameters include switching thresholds, timing parameters and transition conditions. We apply the approach to a laboratory testbed consisting of different production modules which underlines the efficiency improvements for manufacturing units. In addition, plug-and-produce control is enabled by the approach as different configuration of production modules can efficiently be put in operation by re-learning the parameter sets.