Reinforcement Signal Control (RESCO) Benchmark

RESCO provides an interface for traffic network simulation and includes a set of algorithms for traffic signal control.

Quick start

The command below will run the IDQN algorithm on the Cologne Single Signal scenario without using libsumo. Libsumo is faster but requires linux.

cd RESCO/resco_benchmark && pip install -e ../[pfrl,torch]
python main.py @cologne1 @IDQN libsumo:False save_console_log:False gui:True

Features

• Real-world traffic scenarios from three cities: Cologne, Luxembourg, and Salt Lake City.
• Observed and modeled demands from real-world data.
• Salt Lake City scenarios simulate one year of recorded traffic.
• Multiple state, reward, and action space configurations.
• StateBuilder and RewardBuilder interfaces allow for parameterized state and reward functions.
• Three static (non-learning) traffic signal control algorithms: Fixed Time, Max Pressure, and Max Wave.
• Integration with continuous learning benchmark algorithms from COOM.
• Integration with PFRL library algorithms.
• Interfaces with SUMO via subscription API for faster simulation.

Installation

Please visit the installation docs.

Versions

Release v1.0 - RESCO benchmark as of NeurIPS 2021

Release v2.0 - Refactored codebase, more algorithms, saltlake scenarios

Citing RESCO

If you use RESCO in your research, please cite the following paper: Reinforcement Learning Benchmarks for Traffic Signal Control

@inproceedings{ault2021reinforcement,
  title={Reinforcement Learning Benchmarks for Traffic Signal Control},
  author={James Ault and Guni Sharon},
  booktitle={Proceedings of the Thirty-fifth Conference on Neural Information Processing Systems (NeurIPS 2021) Datasets and Benchmarks Track},
  month={December},
  year={2021}
}