Case study
Euro-Information Code Contest
3rd of 177: deciding which AI models to train under an energy budget.
- of 177 participants
- 3rd
- of competition
- 50 h
- datasets in the largest instance
- 250,000
- instances with different constraints
- 7
- 01LP relaxationbound and guide selection
- 02Interval bin packingrealize assignments
- 03CP-SAT · MIPexact sub-problems, repairs
- 04Column generationmost constrained instance
- 05LNS · C/C++improve, hot loops
- 06Distributed searchseveral machines, merge
| Instance | Models | Datasets |
|---|---|---|
| Q1 “Example” | 5 models | |
| Q2 “So it begins.” | 3,000 models | |
| Q3 “Everything is free!” | 5,000 models | |
| Q4 “Those models are demanding!” | 8,000 models | |
| Q5 “Looks like there's a discount on our electricity bill.” | 11,000 models | |
| Q6 “Where are the files?” | 15,000 models | |
| Q7 “We're in the endgame now.” | 50,000 models |
The problem
A data centre must train AI models without exceeding an energy cap. Each model has a value, an energy cost and data requirements per type (numerical, text, image, code), each within a range: below it the model is under-trained, above it energy is wasted.
Each dataset can feed only one model. Using copyrighted data, or data generated by another model, halves the value. A single-requirement model can act as a generator, but a model fed that way cannot feed another.
Seven instances, from 5 models up to 50,000 models and 250,000 datasets. Ranking compares each solution to the best one.
Approach
- LP relaxation over each model's “modes” (clean, degraded, generator) to bound the optimum and guide selection.
- Realization through interval bin packing: fill each requirement with dataset sizes inside its range.
- Exact sub-problems and repairs with CP-SAT (OR-Tools) and MIP (HiGHS, SCIP).
- Column generation with exact dynamic-programming pricing on the most constrained instance.
- Large neighbourhood search (LNS) to improve solutions, with hot loops in C and C++.
- Distributed search: several machines running in parallel, with scripts syncing and merging the best solutions.
Result
3rd of 177 participants. Thanks to Euro-Information for organising it.