An Offline Meta Black-box Optimization Framework for Adaptive Design of Urban Traffic Light Management Systems

Environment Setup

You should install SUMO, an urban traffic simulator. Please follow instructions in here

We strongly recommend to use conda environment. We can install required libraries using requirements.txt

conda create -n tsc python=3.8 -y
conda activate tsc
pip install -r requirements.txt

Main Experiments

Our method consists of three main steps: Offline Data Collection, Meta Training, and Online Adaptation.

• Offline Data Collection

  To collect offline data, you should run the following command. You can set num_worker to accelerate the data collection process via parallelization.

  python collect_data.py --network 2by2 --scheme comb
  

  We provide script file for generating offline data across all networks and schemes in scripts/collect_data.sh

• Meta Training

  After collecting dataset, you should run the following command to train our ANP model. Model will be saved in results/<network>/<scheme>/anp/<exp_id>/ckpt.tar.

  python meta_train.py --network 2by2 --scheme comb --model anp --exp_id trial1
  

  We provide script file for meta-training ANP across all networks and schemes in scripts/meta_train.sh

• Online Adaptation

  Finally, we employ Bayesian optimization with the trained ANP as a surrogate model to find optimal scheme for traffic lights with unseen traffic patterns. You should run the following command to conduct online adaptation.

  python meta_test.py --network 2by2 --scheme comb --model anp --exp_id trial1 --scenario_id 0
  

  We provide script file for online adaptation across all networks and schemes in scripts/meta_test.sh