Test-Fleet Scheduling in Complex Validation and Production Environments

We present a solution to the complex design-automation problem of scheduling test operations in a validation laboratory or production facility. Our goal is to maximize the utilization of a fleet of test stations and minimize the overall test time for a set of products. We consider the realistic scenario where tests can have dependency graphs, implying that some tests must be completed and passed before others can proceed. We also consider a mix of product types that require different kinds of tests and a mix of testers, which implies that each product can only be tested only on a specific set of testers. To ensure scalability and flexibility, we have formulated this scheduling problem as a “partially observable stochastic game”, a multi-agent extension of a partially observable Markov decision process. We have implemented multi-agent reinforcement learning agents to maximize parallelization in a manner that speeds up both training and inferencing. We present scheduling results for synthetic test cases as well as real-life data from a production facility.