Feature Based Parameter/Algorithm Selection Software and Data
Considering that for different problem classes different algorithms may perform better, one problem that arises is how to automatically identify these problem classes and suggest algorithm/parameter setting configurations.
The data folder contains in its subfolder instances problem datasets for different domains. The subfolder instance/mip, for instance, contains a collection of Mixed-Integer Programming instances that were used in experiments of the COIN-OR Branch-and-Cut Solver CBC.
The data subfolder experiments contains experimental results of using different algorithms and parameter settings to solve these problem datasets. The more different algorithms and parameter settings are included in this database the better. Please note that even algorithms and parameter settings that perform poorly in average can be useful for a relevant subset of instances.
Python software is included for the automated generation of decision trees. A decision tree with maximum depth 2 to select parameters for the CBC solver can be generated with the following command:
python dtree.py ../data/instances/mip/features.csv ../data/experiments/cbc/relaxation.csv 2Code is compatible with Pypy just in time compiler, just replace python with pypy if you want it to run faster.
In the previous example, the following decision tree would be generated:
