Differentiable Quadratic Optimization For The Maximum Independent Set Problem

pCQO-MIS v1

ICML 2025

Description

This repository contains the code for the (pCQO-MIS) method. The goal is to provide tools and implementations for the experiments conducted in the submission.

pCQO-MIS v1

pCQO-MIS

Application Setup

Install LibTorch.
Clone the repository and navigate to the ./cpp_impl directory.

Run the following CMake command:

cmake -DCMAKE_PREFIX_PATH={path to libtorch} ..

Build the program using:
```
cmake --build . --config Release
```
The executable pcqo_mis will be available in the ./external directory.

Running the Application

To run the pcqo_mis application, follow these steps:

Build the Application
Ensure the application is built as described in the Application Setup section. The executable pcqo_mis should be available in the ./external directory.
Prepare Input Data
The application requires a graph file in DIMACS format as input. Ensure the graph file is accessible and correctly formatted.
Run the Application
Use the following command to execute the application:
```
./external/pcqo_mis <file_path> [<learning_rate> <momentum> <num_iterations> <num_iterations_per_batch> <gamma> <gamma_prime> <batch_size> <std> <output_interval>] [initialization_vector]
```
- <file_path>: Path to the DIMACS graph file (required).
- All other parameters are optional. If not provided, a grid search will be performed to attempt to find suitable values. Note: This grid search may result in suboptimal parameters that do not yield ideal results. It is strongly recommended to provide these parameters for better performance.
Example Command
For example, to run the application with a graph file graph.dimacs and specific parameters:
```
./external/pcqo_mis graph.dimacs 0.0003 0.875 7500 30 900 1 256 2.25 10
```
For more information about the DIMACS format, see utils/NetworkX_to_DIMACS.ipynb
Output
- The application prints intermediate results at intervals specified by <output_interval>.
- At the end of execution, the maximum independent set found is printed as a binary vector.
Notes
- Ensure the graph file is correctly formatted and accessible.
- Adjust parameters as needed for different graph sizes or optimization requirements.

Benchmark Setup

Prerequisites

Python 3.10
pCQO-MIS Setup
Required libraries: pandas, torch, gurobipy, ortools

Setup and Installation

Clone the repository and navigate to the repository's root directory.
Create a virtual environment using venv:
```
python3 -m venv .venv
```
Activate the virtual environment:
```
source .venv/bin/activate
```
Install the required Python libraries:
```
pip install -r requirements.txt
```
(Optional) If using Gurobi, obtain a license and install it on the machine.
(Optional) If using ReduMIS, clone the KaMIS project, build the ReduMIS program, and place it in the external folder.
Retrieve datasets from the /graphs folder used in the original experiments.
Run the benchmarking suite:
```
python benchmark.py
```

Configuration

Graph Import

Specify the directories containing the graph data by editing the graph_directories list in benchmark.py. For example:

graph_directories = [
    "./graphs/satlib/m403",
    "./graphs/satlib/m411",
    "./graphs/satlib/m418",
]

Solver Configuration

Define the solvers to use in the solvers list. Uncomment the solvers and specify their parameters. For example:

solvers = [
    {
        "name": "Gurobi",
        "class": GurobiMIS,
        "params": {"time_limit": 100},
    },
    {
        "name": "CPSAT",
        "class": CPSATMIS,
        "params": {"time_limit": 30},
    },
]

Warning

When running the pcqo_mis application, ensure that all arguments are provided. If any optional arguments are omitted, the application will default to a slower grid search to determine suitable values. This grid search may result in suboptimal parameters and significantly increase runtime. Providing all arguments is strongly recommended for optimal performance.

Running the Script

Run the script to start the benchmarking process:

python benchmark.py

Checkpoints and Final Results

Intermediate results are saved at intervals defined by SOLUTION_SAVE_INTERVAL. Final results are saved as CSV files named zero_to_stage_{current_stage}_of_{total_stages}_total_stages.csv.

Output

The script generates a CSV file containing results for each graph and solver, including solution sizes and time taken.

Basic Hyper-Parameters Fine-Tuning

For new graphs, a basic hyper-parameter search procedure is provided to assist in setting up $T$ and $\alpha$. You can use it by running pcqo_mis without all optional parameters. Note: This grid search may result in suboptimal parameters that do not yield ideal results. It is strongly recommended to determine these parameters yourself for better performance.

Notes

Ensure graph data and solver implementations are correctly set up and accessible.
Adjust SOLUTION_SAVE_INTERVAL to control the frequency of checkpoint saves.
The benchmarking process may take significant time depending on the number and size of graphs and solvers used.
For large datasets exceeding local RAM, use the benchmark_large_graphs.py script.

ledenmat/pCQO-mis-benchmark