MixedTraffic Benchmark

Official code for the paper:

Optimizing Efficiency of Mixed Traffic through Reinforcement Learning: A Topology-Independent Approach and Benchmark

Chuyang Xiao, Dawei Wang, Xinzheng Tang, Jia Pan, Yuexin Ma

Environment Set-up

Requirements -- Anaconda Set-up

We highly recommend installing on Ubuntu 22.04 as this version of Ubuntu has been tested. If you have a Windows machine, we recommend using WSL to create an Ubuntu 22.04 virtual machine for setting up the repo.

ubuntu

conda create -n MixedTrafficPlus python=3.8
conda activate MixedTrafficPlus
conda env update -f MixedTrafficPlus.yml

Install SUMO

sudo apt-get install sumo sumo-tools sumo-doc

Training

• To train on various topologies which involve intersections and roundabouts, you can run the following command and choose the rv_rate you want to use as the folder name.

--resume_dir default="checkpoint". -> It will automatically find the latest checkpoint in this checkpoint folder
--wandb-id default="". -> It will automatically create a new wandb run witha new wandb_id

cd benchmark_training
cd large_rv={rv_rate}
python train_large_benchmark.py --resume-dir="the/path/to/the/checkpoint/you/want/to/resume" --wandb-id "wandb-id_you_want_to_use"

• To train on only one topologies, for example, roundabouts, you can run the following command with rv_rate=1.

cd our_roundabout
python train_our_ra.py --resume-dir="the/path/to/the/checkpoint/you/want/to/resume" --wandb-id "wandb-id_you_want_to_use"

Testing

• To test on various topologies which involve intersections and roundabouts, you can run the following command and choose the rv_rate you want to use as the folder name.

cd benchmark_training
cd large_rv={rv_rate}
python test_large_benchmark.py --resume-dir="the/path/to/the/checkpoint/you/want/to/resume" --wandb-id "wandb-id_you_want_to_use"

• To test on only one topologies, for example, roundabouts, you can run the following command with rv_rate=1.

cd our_roundabout
python test_our_ra.py --resume-dir="the/path/to/the/checkpoint/you/want/to/resume" --wandb-id "wandb-id_you_want_to_use"

Model Checkpoints

Pre-trained model checkpoints are available on Hugging Face Hub:

• MixedTrafficPlus Repository
• RV=100% Model Checkpoint (trained with our method)

Generate More Data

You can put your osm files downloaded from openstreetmap.org to train_file or test_file folder and select the folder difficulty(mode = 400 veh/h, easy(1000 veh/h), middle(3000 veh/h), hard(5000 veh/h)). Then you can run

./scripts/generate_train_set/generate_{mode}_routes.sh 

if you want to generate training scenario files.