Hyperparameter Optimization for SCIP using Hyperband

GRIPS 2022

Students: M. Aïdli, B. Liang, E. Vercesi, A. Zhang

Instruction

This repository contains basic script to asses the performances of Hyperband, SMAC and GGA

Set up the `conda` environmrnt

In the main folder, run

 conda env create -f conda.yaml --name ml4co_grips #Replace ml4co_grips with your favourite name
 conda deactivate # Run only if you have any conda environment active
 conda activate ml4co_grips

Instances

Here you can find all the information for downloading the dataset. Note that the dataset should go in the ./instances folder

Test Hyperband

Move to the folder hb and run

python main_scip.py

Test SMAC

Move to the folder SMAC and run

python main_smacwithscip.py

Test GGA

To the GGA, you firstly need a working version of OPTANO Then, you can go in the folder GGA and run

./runOptano.sh

Other codes

We have also implemented a per-instance approach that strongly relies on feature extraction. To extract feature from a bunch of instances you can use the script get_instance_feature.py

Tested on

Debian GNU/Linux 11 (bullseye)

Note A large portion of the Hyperband implementation was adapted from https://github.com/zygmuntz/hyperband.

Contacts

Mélissa Aïdli : melissa.aidli@student-cs.fr

Eleonora Vercesi : eleonora.vercesi01@universitadipavia.it

Annie Zhang : annie.zhang@uvm.edu

biyonka/ml4co_grips