Minerva: A reinforcement learning-based technique for optimal scheduling and bottleneck detection in distributed factory operations

In manufacturing systems, the term bottleneck refers to a component that limits the entire throughput of a system. A number of approaches have attempted bottleneck detection. However, existing solutions have limitations, leaving the bottleneck identification still no trivial task. To address the shortcomings of prior works, we study Job Shop Scheduling Problems (JSSP) with the realistic extension that jobs are enqueued periodically, and propose a machine learning based solution to such a problem, named Minerva. Minerva first finds the optimal resource scheduling for a target interval, based on a model-free reinforcement learning technique. Then, using an artificial neural network classifier, Minerva identifies the constrained resources for each target interval. Minerva is evaluated on two representative benchmarks with the key result being that Minerva is able to detect the system bottleneck(s) with a high accuracy of 95.2%, which is almost 25% better than the best-in-class bottleneck identification methods.