2246. Linearizing Convolutional Neural Networks for Performance Evaluation and Optimization of Stochastic Flow Lines
Invited abstract in session WE-10: Flow Lines and Fairness Considerations, stream Supply Chain Management and Production.
Wednesday, 16:30-18:00Room: H16
Authors (first author is the speaker)
| 1. | Andreas Kloos
|
| Chair of Production Management, Universität Regensburg | |
| 2. | Justus Arne Schwarz
|
| Chair of Production Management, University of Mannheim |
Abstract
Simulation is widely used for evaluating stochastic manufacturing systems, due to its adaptable modelling capacities and the absence of exact or even approximate analytical solutions. Solving decision problems with integrated simulation models can be time-consuming, due to the need for many replications and the combinatorial nature of many problems. To increase the speed of the evaluation of the performance measures of a manufacturing system, the existing literature suggests replacing simulation with Artificial Neural Networks pretrained with simulation results. We train a Convolutional Neural Network (CNN) with flowline feature data to predict the throughput of a flowline configuration. We then solve optimization problems, by linearizing the trained CNN, thereby transforming the non-linear problem into a mixed-integer linear problem (MILP), which is tractable by standard solvers. Furthermore, the impacts of different sampling methods for generating training data and regularization methods are investigated.
Keywords
- Machine Learning
- Mixed-Integer Programming
- Production and Inventory Systems
Status: accepted
Back to the list of papers