Neural Focused Ant Colony Optimization (NeuFACO) is a non-autoregressive framework for solving the Traveling Salesman Problem (TSP). It combines reinforcement learning with advanced Ant Colony Optimization (ACO) techniques to achieve both efficiency and solution quality. The code is built upon the codebase of GFACS, which itself is developed based on the DeepACO codebase.
Uses Proximal Policy Optimization (PPO) to train a Graph Neural Network (GNN).
Generates instance-specific heuristic guidance rather than relying on fixed training schemes.
Incorporates candidate lists, focused tour modification, and scalable local search.
Efficiently integrates the learned heuristic into the solution process.
Leverages amortized inference together with the parallel stochastic exploration of ACO.
Provides fast and high-quality solutions across diverse TSP instances.
- Python 3.11.5
- PyTorch 2.1.1
- PyTorch Geometric 2.4.0
We strongly recommend using uv for virtual environment in this project.
uv venv python=3.11
uv pip install torch torch_geometric
uv pip install numpy numba pandas pyvrp scipy tdqm wandb
For the complete list of dependencies, please refer to the requirements.txt file.
For the usage of the code, please refer to each folder's README.md file.