MacLight

Code of AAMAS2025 full paper "MacLight: Multi-scene Aggregation Convolutional Learning for Traffic Signal Control".

Package Requirements

Python 3.8, torch, numpy, pandas, tqdm, eclipse-sumo, traci, libsumo, sumo-rl

Except for torch, the other packages can be installed by running the commands in the install.sh script. Torch needs to be installed separately.

Environment Setup

1. Use conda to create a virtual environment, specifying python=3.8. Let's assume the environment name is your_env.

2. Install torch following the instructions on the official Torch website.

3. Use conda or pip to install is ok for

   pip install eclipse-sumo
   pip install traci
   pip install libsumo
   pip install sumo-rl
   pip install gymnasium
   pip install pettingzoo
   

4. Set the SUMO path. On Linux:

   export SUMO_HOME=miniconda3/envs/{your_env}/lib/{python3.8}/site-packages/sumo

   (Where miniconda might be anaconda). On Windows, set an environment variable pointing to the SUMO folder.

Run MacLight

To run MacLight:

python run_Ours.py -w 1 -t block -l normal

After the execution is complete, the data will be saved in the data/plot_data/ folder.

Test your method

Experimental Limitations

If you want to transfer this method, you can adjust the road xml file in the SUMO interface in run_{algorithm}.py. But it cannot be transferred arbitrarily because we have specified the data organization, that is, the road network needs to be a grid.

You can also refer to our method for constructing dynamic traffic flow and our method for constructing variational autoencoders.

Dynamic traffic flow: env\wrap\random_block.py. We also give a guide on how to build it, see the subsequent section Dynamic Traffic Flow Construction Guide.

variational autoencoders: net\net.py.

Train Your Method in Our Road Network

Road network files

If you plan to test your method on our road network, you can find road network files in env\map.

• ff.net.xml is road network. It is unchanging, and the difference between the different environments is the traffic flow file.
• ff_normal.rou.xml is traffic flow file under normal pressure (Normal&Block).
• ff_hard.rou.xml is traffic flow file under high pressure (Peak).

It should be noted that the normal pressure we set is also relatively high.

Traffic flow reference table:

example to create an enviroment

Our environment interface inherits from gymnasium and sumo-rl, so you can easily migrate your algorithms.

1. Static environment:

   The direction of the vehicle is completely fixed and the route will not be changed.

   env = sumo_rl.parallel_env(net_file='env/map/ff.net.xml',
                              route_file=f'env/map/ff_normal.rou.xml',  # Could be `ff_hard.rou.xml`
                              num_seconds=args.seconds,
                              use_gui=False,
                              sumo_warnings=False,
                              additional_sumo_cmd='--no-step-log')

2. Dynamic environment

   We define the BlockStreet class, which can randomly block certain roads, so that vehicles reselect the best route, which will cause sudden changes in traffic flow on certain roads.

   from env.wrap.random_block import BlockStreet
   # block_num: Number of blocked roads, like 8
   # seconds: Simulation seconds, up to 3600
   env = BlockStreet(env, block_num, seconds)

Additionally, we strongly recommend that you set the following environment variables to get the fastest possible simulation (although it may still be slower):

import os
os.environ['LIBSUMO_AS_TRACI'] = '1'

Full Experimental Result

A complete presentation of some experimental figures in the paper. Also gives the parameter settings.

Parameters

Road blocking experiment statistics

Complete statistics on the changes in the distribution of traffic flow before and after the implementation of road blockage. Road map reference paper or subsequent section.

Full training

Dynamic Traffic Flow Construction Guide

Block Environment Code

Refer to the file env\wrap\random_block.py.

Simulation Environment Creation

After installing SUMO and configuring its environment variables, you can create the simulation environment.

1. Use the following command to create a simple road network:

   netgenerate --grid --grid.number=6 --grid.length=200 --default.lanenumber=6 -o ff.net.xml

2. In netedit, manually delete the surrounding roads and the corner nodes.

3. Create traffic lights:

   netconvert --sumo-net-file ff.net.xml --tls.guess --output-file ff.net.xml

4. Delete the left-turn and straight signals for the left lanes of the traffic lights.

5. Modify the traffic light phases in the ff.net.xml file (do not use the default phases) by locating the </tlLogic> tag and adjusting all traffic light phases to:

   <phase duration="30" state="grrrgGGrgrrrgGGr"/>
   <phase duration="3"  state="grrrgGGrgrrrgyyr"/>
   <phase duration="15" state="grrrgGGGgrrrgrrr"/>
   <phase duration="3"  state="grrrgyyygrrrgrrr"/>
   <phase duration="15" state="grrrgrrrgrrrgGGG"/>
   <phase duration="3"  state="grrrgrrrgrrrgyyG"/>
   <phase duration="15" state="grrrgrrGgrrrgrrG"/>
   <phase duration="3"  state="grrrgrrygrrrgrry"/>
   <phase duration="30" state="gGGrgrrrgGGrgrrr"/>
   <phase duration="3"  state="gyyrgrrrgGGrgrrr"/>
   <phase duration="15" state="grrrgrrrgGGGgrrr"/>
   <phase duration="3"  state="grrrgrrrgyyygrrr"/>

Vehicle ID: 0–8999

To get the list of vehicle IDs currently operating in the system:

traci.vehicle.getIDList()

To reroute based on travel time:

traci.vehicle.rerouteTraveltime(vehicle_id)

Map

Blocking Roads

To block a road by reducing its speed limit:

traci.edge.setMaxSpeed(edge, 1)

ID List of All Lanes

['A1B1', 'A2B2', 'A3B3', 'A4B4',
 'B0B1', 'B1A1', 'B1B0', 'B1B2',
 'B1C1', 'B2A2', 'B2B1', 'B2B3',
 'B2C2', 'B3A3', 'B3B2', 'B3B4',
 'B3C3', 'B4A4', 'B4B3', 'B4B5',
 'B4C4', 'B5B4', 'C0C1', 'C1B1',
 'C1C0', 'C1C2', 'C1D1', 'C2B2',
 'C2C1', 'C2C3', 'C2D2', 'C3B3',
 'C3C2', 'C3C4', 'C3D3', 'C4B4',
 'C4C3', 'C4C5', 'C4D4', 'C5C4',
 'D0D1', 'D1C1', 'D1D0', 'D1D2',
 'D1E1', 'D2C2', 'D2D1', 'D2D3',
 'D2E2', 'D3C3', 'D3D2', 'D3D4',
 'D3E3', 'D4C4', 'D4D3', 'D4D5',
 'D4E4', 'D5D4', 'E0E1', 'E1D1',
 'E1E0', 'E1E2', 'E1F1', 'E2D2',
 'E2E1', 'E2E3', 'E2F2', 'E3D3',
 'E3E2', 'E3E4', 'E3F3', 'E4D4',
 'E4E3', 'E4E5', 'E4F4', 'E5E4',
 'F1E1', 'F2E2', 'F3E3', 'F4E4']

Restricting the Central Area as a Starting Point

Refer to the ff.dst.xml and ff.src.xml files to ensure that roads in the central area are not used as starting points. Use the corresponding parameter when generating trips with randomTrips.py: --weights-prefix ff (no suffix is required).

Allowed starting edges:

["A1B1", "B1A1", "A2B2", "B2A2", "A3B3",
"B3A3", "B4A4", "A4B4", "E1F1", "F1E1",
"E4F4", "F4E4", "E2F2", "F2E2", "E3F3",
"F3E3", "B1B0", "B0B1", "C1C0", "C0C1",
"D1D0", "D0D1", "E1E0", "E0E1", "B5B4",
"B4B5", "C5C4", "C4C5", "D5D4", "D4D5",
"E5E4", "E4E5"]

Prohibited starting edges:

['B1B2', 'B1C1', 'B2B1', 'B2B3', 'B2C2',
 'B3B2', 'B3B4', 'B3C3', 'B4B3', 'B4B5',
 'B4C4', 'B5B4', 'C1B1', 'C1C2', 'C1D1',
 'C2B2', 'C2C1', 'C2C3', 'C2D2', 'C3B3',
 'C3C2', 'C3C4', 'C3D3', 'C4B4', 'C4C3',
 'C4D4', 'D1C1', 'D1D2', 'D1E1', 'D2C2',
 'D2D1', 'D2D3', 'D2E2', 'D3C3', 'D3D2',
 'D3D4', 'D3E3', 'D4C4', 'D4D3', 'D4E4',
 'E1D1', 'E1E2', 'E2D2', 'E2E1', 'E2E3',
 'E3D3', 'E3E2', 'E3E4', 'E4D4', 'E4E3']

These edges may be blocked, so they cannot be set as start or end points.

Cite

If our repository is helpful to you, you can cite it in your research.

@misc{MacLight2024Lee,
    author = {Sunbowen Lee},
    title = {{MacLight}},
    year = {2024},
    publisher = {GitHub},
    journal = {GitHub repository},
    howpublished = {\url{https://github.com/Aegis1863/MacLight}},
}

We also recommend looking at sumo-rl, our environment is based on their implementation.