2218. Embedding Neural Networks into Optimization Models with GAMSPy
Invited abstract in session WE-9: Modelling and APIs, stream Software for Operations Research.
Wednesday, 16:30-18:00Room: H15
Authors (first author is the speaker)
| 1. | Frederik Fiand
|
| GAMS Software GmbH | |
| 2. | Michael Bussieck
|
| GAMS Software GmbH | |
| 3. | Hamdi Burak Usul
|
Abstract
GAMSPy is a powerful mathematical optimization package which integrates Python's flexibility with GAMS's modeling performance. Python features many widely used packages to specify, train, and use machine learning (ML) models like neural networks. GAMSPy bridges the gap between ML and conventional mathematical modeling by providing helper classes for many commonly used neural network layer formulations and activation functions. These allow a compact description of the network architecture that gets automatically reformulated into model expressions for the GAMSPy model.
In this talk, we demonstrate how GAMSPy can seamlessly embed a pretrained neural network into an optimization model. We also explore the utility of GAMSPy's automated reformulations for neural networks in various applications, such as adversarial input generation, model verification, customized training, and leveraging predictive capabilities within optimization models.
Keywords
- Modeling Systems and Languages
- Machine Learning
Status: accepted
Back to the list of papers