ContrArc-MILP: Learning-Oriented Binary Integer Programming Dataset

A dataset of 2,000 binary integer linear programming (BILP) instances derived from contract-based architecture selection problems (ContrArc). Designed for training and evaluating GNN-based predict-and-search solvers with controlled distribution shifts.

Dataset Structure

Two problem domains, each with four splits:

Split	EPN	RPL	Description
train	600	600	Training set (default constraints, moderate size)
test_id	150	150	In-distribution test (same distribution as train)
test_large	100	100	Scale shift (~2x larger instances, same constraints)
test_ood	150	150	Structural shift (tighter coupling constraints)

Cross-domain generalization can be evaluated by training on one domain and testing on the other (e.g., train on epn/train, evaluate on rpl/test_id).

Domains

• EPN (Electric Power Network): Architecture selection for aircraft power distribution networks with generators, AC/DC buses, rectifiers, and loads.
• RPL (Reconfigurable Production Line): Configuration of reconfigurable manufacturing production lines with conveyors and machines.

Instance Statistics

Split	N	Median vars	Median constraints	Median SCIP time	Feasibility	Optimality
epn/train	600	13,308	12,740	3.0s	100%	100%
epn/test_id	150	13,328	12,884	4.7s	100%	100%
epn/test_large	100	27,072	19,832	11.9s	100%	100%
epn/test_ood	150	14,044	13,308	4.8s	100%	100%
rpl/train	600	14,990	10,480	2.2s	100%	92%
rpl/test_id	150	14,794	10,480	2.8s	100%	91%
rpl/test_large	100	33,540	16,028	4.5s	93%	82%
rpl/test_ood	150	16,656	11,580	3.2s	100%	97%

SCIP times measured with a 300-second time limit. Optimality = gap < 1e-6.

File Format

Each instance is stored as a compressed NumPy archive (.npz):

import numpy as np
import gzip

data = np.load("epn/train/00000.npz", allow_pickle=True)

# Gzip-compressed LP file content (CPLEX LP format)
lp_string = gzip.decompress(bytes(data["lp_gz"])).decode()

# Optimal binary solution vector from Gurobi
solution = data["solution"]  # shape: (n_vars,), dtype: float32, values in {0, 1}

# Optimal objective value
obj_val = float(data["obj_val"])

Each split directory also contains:

• manifest.json: Per-instance metadata (n_vars, n_constrs, obj_val)
• results.json: SCIP solve results (solve time, gap, feasibility)

Distribution Shift Design

The dataset tests three types of generalization:

1. In-distribution (test_id): Same component ranges and constraint rules as training. Baseline for model performance.

2. Scale generalization (test_large): ~1.5-2x larger instances with the same constraint structure. Tests whether learned heuristics transfer to bigger problems.

3. Structural generalization (test_ood): Same instance sizes but with tighter coupling constraints (cross-tag composition rules, implication constraints). Tests robustness to constraint distribution shift.

4. Cross-domain generalization (implicit): Train on EPN, evaluate on RPL (or vice versa). The domains have different component types and connectivity patterns.

Problem Characteristics

All instances are pure binary (0/1) integer programs with:

• Linear objective (minimization)
• Linear constraints (equalities and inequalities)
• Derived from Gurobi models with AND, OR, and indicator general constraints, linearized via big-M formulation

The problems encode architecture selection: choosing components, their implementations, and connection mappings to minimize total cost while satisfying structural and contractual constraints.

Intended Use

This dataset is designed for:

• Training GNN-based MILP solvers (e.g., predict-and-search, Neural Diving)
• Evaluating generalization of learned combinatorial optimization methods
• Benchmarking under controlled distribution shifts

Generation

Instances are generated using ContrArc, a contract-based methodology for cyber-physical system architecture exploration that converts architecture selection into MILP using assume-guarantee contracts and subgraph isomorphism (Xiao et al., DATE 2024). Instances are generated with randomized component counts, library sizes, and constraint configurations. Optimal solutions are obtained via Gurobi. SCIP solve times are provided as reference solver baselines.

Citation

If you use this dataset, please cite:

@inproceedings{xiao2024contrarc,
  title={Efficient Exploration of Cyber-Physical System Architectures Using Contracts and Subgraph Isomorphism},
  author={Xiao, Yifeng and Oh, Chanwook and Lora, Michele and Nuzzo, Pierluigi},
  booktitle={2024 Design, Automation \& Test in Europe Conference \& Exhibition (DATE)},
  pages={1--6},
  year={2024},
  doi={10.23919/DATE58400.2024.10546764}
}