Machine Learning for Combinatorial Optimization - NeurIPS 2021 Competition (Configuration task)

This is the code repository of team "MixedInspiredLamePuns"

Note that this repository is forked from github.com/ds4dm/ml4co-competition. See here for the official competition website.

For a related application, trying to explore the dataset using Bayesian Optimization, check out the bo_gnn_seminar branch.

Leaderboard results

• rank 3 / 15 on the global leaderboard
• rank 1 / 7 on the student leaderboard

Arxiv Publication

Read our short write-up.

Poster

See also the pdf version.

Code structure

.
├── bo_gnn                              <-- main directory
│   └── config
│       ├── conda.yaml
│       ├── config_utils.py
│       ├── data_utils
│       │   ├── dataset.py              <-- dataset loader
│       │   └── milp_data.py            <-- bipartite graph representation
│       ├── Dockerfile.cpu              <-- Dockefile for data generation
│       ├── Dockerfile.gpu              <-- Dockerfile for GNN training
│       ├── generate_data.py            <-- massively parallel data generation script
│       ├── models
│       │   ├── baseline.py             <-- GNN definition
│       │   └── callbacks.py
│       ├── mps_to_pkl.py
│       ├── plot_run_performance.py
│       ├── scripts
│       └── train_gnn.py                <-- main training loop & ensembling
├── common                              <-- pre-defined evaluation environments
│   ├── environments.py
│   ├── evaluate.py
│   └── rewards.py
├── instances                           <-- raw data location
├── README.md
├── submissions                         <-- submission code
│   └── MixedInspiringLamePuns
│       ├── agents
│       ├── callbacks.py
│       ├── conda.yaml
│       ├── data_utils
│       ├── heuristics_schedules
│       ├── init.sh
│       ├── models
│       ├── param_configurations
│       ├── parameter_configuration_mapping
│       ├── results
│       ├── trained_model_checkpoints
│       └── train_gnn.py
└── TASKS.md