[TOC]

Cross-Problem Solving for Network Optimization: Is Problem-Aware Learning the Key?

This repository is the official repository for the paper Cross-Problem Solving for Network Optimization: Is Problem-Aware Learning the Key? (Accepted by IEEE JSAC).

Setup

conda env create -f environment.yml
conda activate PAD

Codebase Structure

datatool/ includes problem formulations, dataset creation tools, dataset preprocessing tools, and problem feature embedding creation and analysis tools.

diffusion/ includes encoder-decoder model, constraint-aware module, PAD model.

runner/ includes training and testing files of encoder-decoder model, constraint-aware module, PAD model.

scripts/ includes scripts that call each run file in runner/.

Reproduction

If you want to add new problems, you can add them in the problem_settings.py and diff_train_test_cfg.json according to the specified format.

# Prepare datasets
# If new problems are added, the corresponding data producing tools and related data processing functions need to be implemented
# in data_producing.py, data_preprocess.py, solution_process.py
cd datatool/
python data_producing.py
# Before this, please download the corresponding MathBert or DeepSeek model to your local machine
# Generate and store all problem embeddings
python formula_embedding.py
# Train and test
cd scripts/
./enc_dec.sh
# Add the trained encoder-decoder model path to the script and run
./cons_module.sh
# Add the trained encoder-decoder model path and constraint-aware module path to the script and run
./pad.sh

Data and Pretrained Checkpoints

Please refer to the paper for detailed data introduction. If you need data files or the results (ckpts, etc.) in the paper, please contact us by email.