This code was written as part of our bachelor thesis titled: "Reinforcement learning for improved local search, applied to the graph coloring problem". This work was done at the KTH Royal Institute of Technology in Stockholm, Sweden. This repository contains the code for the RLTCol algorithm, which is a hybrid heuristic algorithm for the graph coloring problem that uses reinforcement learning (RL).
The RLTCol algorithm works by iteratively running the local search algorithm TabuCol, and running an RL agent. The two components pass solutions to each other. The paper can be read here.
The RL agent is implemented in Python using the Tianshou library. TabuCol is implemented in Rust, using maturin to interface with Python. The code is written for Python 3.10.
- Python 3.10
- A working Rust installation, see here for instructions.
Create a virtual environment and install the required Python packages:
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txtThen, build the Rust code:
cd src/tabucol && maturin develop --release && cd -The graphs used as input need to be in the form of a DIMACS text file. The graphs used in the paper can be found here. If they are in the compressed format, they can be decompressed using the translator found on the same page.
The source code for the RLTCol algorithm is located in the src directory. The TabuCol implementation in Rust is located in the src/tabucol directory.
The RL agent can be trained using the trainer.py script. The script takes a number of arguments, which can be found by running python trainer.py --help.
The script will save the trained policy to the file specified by the output parameter. The policy can then be used to run the RLTCol algorithm.
The RLTCol algorithm can be run using the runner.py script. The script takes a number of arguments, which can be found by running python runner.py --help.
In order to run multiple jobs in parallel and/or in sequence, the batch_runner.py script can be used. The script takes a number of arguments, which can be found by running python batch_runner.py --help. This script will run the RLTCol algorithm and save the results to individual files in the directory specified by the output_dir parameter. The results can then be summarized using the result_summarizer.py script.
The paper was written by Adrian Salamon (asalamon@kth.se) and Klara Sandström (klarasan@kth.se), supervised by Stefano Markidis.