code for paper PlanLight: Learning to Optimize Traffic Signal Control with Planning and Iterative Policy Improvement.
An RL and planning based method for TSC simulated on CityFLow.
train PlanLight from scratch
python run_iterations.py CONFIG_FILE_PATH
The config file is required for running CityFlow, we provided several sample configs in ./config/
Setting existing method as the first base_policy for PlanLight
-
train the model you want to use. eg. CoLight:
python run_colight.py CONFIG_FILE_PATH --save_dir MODEL_PATHthe first step can be ommited using non-RL-based methods as first base policy
-
run onestep rollout to collect trajectories:
python run_iteration.py CONFIG_FILE_PATH --model_dir MODEL_PATH --base_policy colight -
run PlanLight
python run_iterations.py CONFIG_FILE_PATH --head_start_traj_name TRAJECTORY_FILE