Adaptive Traffic Signal Control on Devices with Extremely Limited Resources

Code for Adaptive Traffic Signal Control on Devices with Extremely Limited Resources (IJCAI 2022).

Dependencies

python=3.6.13
torch=1.9.0+cu111
numpy=1.19.2
CityFlow=0.1.0

We made some modifications based on CityFlow v0.1.0 to enrich its APIs. To ensure reproducibility, please install our adapted CityFlow by:

cd CityFlow
pip install .

Experiments on ATSC

Note: all the following steps are performed in TinyLight folder.

STEP 0: modify `config.json` (optional)

STEP 1: run baseline

python 01_run_baseline.py --model=FixedTime --dataset=Atlanta

STEP 2: sub-graph extraction

The model will be stored in {log_path}/TinyLight/model/.

python 02_run_tiny_light.py --dataset=Atlanta

STEP 3: post-training quantization

This step converts the floating-point operations into integer-only ones. It will generate a series of model files in {log_path}/TinyLightQuan/pc and {log_path}/TinyLightQuan/mcu, which are used for evaluation in step 4.

python 03_run_tinylight_quan.py --dataset=Atlanta

STEP 4: evaluate TinyLight (w/ PTQ)

You can run TinyLight on an Arduino Uno (with ATmega328P as the default MCU) by:

replacing TinyLight_MCU/model.h with the corresponding .h file in {log_path}/TinyLightQuan/mcu, and
running the project in MCU

If you do not have an Arduino board on hand, we also provide an equivalent implementation of TinyLight (w/ PTQ) on computers to verify its performance on traffic. For this, please:

cd {log_path}/TinyLightQuan/mcu/ 
python setup.py build_ext --inplace
cd {project_dir}
python 04_run_tinylight_mcu.py --dataset=Atlanta

BibTeX

If you find our work helpful in your research, please consider citing our paper:

@article{TinyLight22,
  author    = {Dong Xing and
               Qian Zheng and
               Qianhui Liu and
               Gang Pan},
  title     = {TinyLight: Adaptive Traffic Signal Control on Devices with Extremely Limited Resources},
  journal   = {CoRR},
  volume    = {abs/2205.00427},
  year      = {2022},
  url       = {https://doi.org/10.48550/arXiv.2205.00427},
  doi       = {10.48550/arXiv.2205.00427},
  eprinttype = {arXiv},
  eprint    = {2205.00427},
}

yuanzhi0515/TinyLight