The ``Factory Coating Problem'' is a commercially important problem which several members of our lab group are working on. It involves a huge production line that produces about 100 products per second which need to be coated by a special process. Depending on the desired final product, there are about twenty different possible coating processes. The line can be configured with a small number of different coaters. Each coater, when it is not busy and when it is working properly, greedily grabs products from the line. If too many coaters are busy, or too many are temporarily out of service, products reach the end of the line and (literally) fall into an enormous waste bin.
The problem is to find a good schedule for a month of production. The start state is a list of demands for how many of each product must be manufactured (numbers are typically in the tens or hundreds of millions). The task is to find a sequence of configurations over the month to meet production requirements. The objective function combines minimizing wasted products, minimizing the number of configuration changes, and completing the month's production as quickly as possible. Given an approximate model of the performance of individual coaters and their conditional probabilities of temporary breakdowns, the problem can be specified as an MDP with a state space formed as the Cartesian product of { a 20-dimensional demand space } { approximately 85,000 coater configurations } {time to end of month}. This MDP is acyclic and thus makes an excellent practical application for ROUT.