Algorithm selection addresses the problem of automatically determining the best algorithm for a given task (often called instance or input to the algorithm). A well known approach is to train a machine learning model which predicts for a given instance which algorithm to use. Previous studies showed that this can lead to automatic performance improvements between factor 4 and 10. Although there exist many different machine learning approaches to solve the algorithm selection problem, how to use deep learning in this context is rarely studied. The task in this project is to (i) use AutoML approaches to find a well-suited architecture for different algorithm selection datasets, (ii) study different loss functions for algorithm selection and (iii) run large scale experiments on ASlib. The project will be based on AutoFolio and its implementation.
We strongly recommend that students should have hands-on experience for the following topics:
Contact person: Marius Lindauer