Princeton Consultants, a simulation consulting firm, serves a manufacturing client with a hard machine scheduling problem. Its optimizer had difficulty scheduling machines for new types of items that needed to be processed; it was not able to respond quickly and accurately to change. A trained Pathmind policy was able to respond to new items without crashing, and while being updated quickly, in order to maximize the number of items they can process before hitting a time limit.
Like most manufacturers, Princeton Consultants’ client needed to process a steady stream of new items in its plant, and to decide which machines should do the processing on the fly.
The manufacturer’s Mixed Integer Programming (MIP)-based optimizer crashed frequently, because it was unable to handle the variability of input when new items were introduced to the line. Adjusting the optimizer required writing hundreds of lines of code, which could take weeks.
In sum, Princeton Consultants needed to build an optimizer that kept up with new customer requests, while saving on cost incurred by expensive optimizers. The manufacturing shop sought to make faster and cheaper decisions – without loss to performance.
Why Pathmind Was Needed
Using a Pathmind reinforcement learning policy, Princeton Consultants was able to produce an adaptive solution that could optimize machine scheduling for new items without crashing, and which could be updated in less than an hour. Normally, adapting the previous optimizer to new items was laborious and time consuming.
Reinforcement learning allows simulation modelers to train on a wide range of possibilities, which can result in a policy that generalizes well to new situations. In addition, reinforcement learning policies can be retrained on new data in a matter of hours, without needing to rewrite any code.
Princeton Consultants was able to produce a Pathmind policy that could handle more new items without crashing, which increased the number of items successfully processed. It kept up with new customer requests, without requiring a rewrite of the optimizer.
It did so by reducing the number of products that expired before they could be processed (imagine a meat-packing plant, for example, where the goods are perishable). In addition, the simulation modeler did not need to spend weeks rewriting optimizer code. (As an added bonus, the Pathmind Policy beat the previous heuristic by 2%, and offers the potential for greater gains through further reward shaping.)
About Princeton Consultants
With more than 35 years experience, Princeton Consultants have completed over 1,500 projects for many of the most successful and innovative companies in the world, such as FedEx, Allstate, Pfizer, Panasonic and GE. Princeton Consultants help clients transform their performance by integrating new technologies into their core business process.
Contact us to learn more about this machine scheduling use case and other Pathmind projects.