search

ds4dm / ecole

Extensible Combinatorial Optimization Learning Environments
https://www.ecole.ai
BSD 3-Clause "New" or "Revised" License
319 stars 68 forks source link

Ecole instance generators don't match learn2branch generators #273

Open dchetelat opened 2 years ago

dchetelat commented 2 years ago

Discussed in https://github.com/ds4dm/ecole/discussions/272

Set covering problems and maximum independent set problems look too easy, and the combinatorial auctions and capacitated facility locations look too hard.

dchetelat commented 2 years ago

Additional details: I ran a side-by-side experiment on setcover instances. (Code can be found here.) 500 instances were generated both from the original learn2branch generators and the Ecole generators. They are solved with the same Ecole Configuring environment, same SCIP version, with the same hyperparameters. The resulting number of nodes were the following (1-shifted geometric mean):

learn2branch generators : 47.85 nodes
Ecole generators        : 4.65 nodes

So there seems to be indeed a discrepancy in difficulty between the old and new setcover instances. Presumably the same holds with the other families.