RealLight is a reinforcement learning framework for traffic signal control under observations limited to what an intersection surveillance camera can provide.
Each intersection is controlled by an independent agent. The agent observes lane-level vehicle densities within a 100 m range together with aggregate flow indicators for neighboring intersections, selects a joint phase and green duration, and is trained with independent Proximal Policy Optimization (IPPO). The state and the reward are computed from aggregate counts inside a prescribed observation range rather than from exact trajectories, unrestricted queues, or network-wide state supplied by the simulator.
Because each agent has its own observation and action dimensions, the formulation supports three-way, four-way, and multi-leg intersections with different lane counts and intersection-specific phase sets, without padding or parameter sharing.
- Python 3.8+
- CityFlow
- PyTorch 2.1.0+
- numpy, pyyaml, pandas
git clone https://github.com/camuslab/RealLight.git
cd RealLight
python run.pyHyperparameters and the scenario to run are set in conf.yaml.
| Path | Contents |
|---|---|
run.py |
Training and evaluation entry point |
env.py |
CityFlow environment wrapper |
conf.yaml |
Hyperparameters and scenario selection |
algos/reallight.py |
RealLight (IPPO) agent |
algos/fixed_time.py |
Fixed-time baseline |
data/synthetic/ |
1x3, 2x2, 3x3, and 4x4 grid scenarios |
data/real/ |
Jinan, Hangzhou, New York, and Daejeon Seo-gu scenarios |
The Daejeon Seo-gu network is newly constructed for this work. Its geometry is built from OpenStreetMap data, and its signal phases follow existing time-of-day operational plans. It contains three-way, four-way, and multi-leg intersections, one to five lanes per approach, and shared straight-left and straight-right lanes.
Average travel time in seconds. RealLight-100 uses the 100 m observation
range; RealLight-max removes the range limit.
| Method | 1x3 | 2x2 | 3x3 | 4x4 | Jinan | Hangzhou | New York | Seo-gu | |
|---|---|---|---|---|---|---|---|---|---|
| Non-RL | Fixed-Time | 384.47 | 454.01 | 508.87 | 565.99 | 405.91 | 488.51 | - | 220.56 |
| SOTL | 247.07 | 331.64 | 424.66 | 474.32 | 410.65 | 505.53 | - | - | |
| RL | GRL | 208.21 | 239.13 | 431.43 | 523.01 | 562.91 | 598.17 | - | - |
| CoLight | 210.01 | 312.29 | 328.70 | 397.07 | 327.62 | 337.45 | 1459.28 | - | |
| PressLight | 98.74 | 123.90 | 166.28 | 215.32 | 285.65 | 341.99 | - | - | |
| IPDALight | 88.01 | 109.66 | 146.92 | 184.54 | 255.35 | 298.99 | - | - | |
| Ours | RealLight-100 | 85.71 | 107.08 | 142.83 | 181.39 | 253.39 | 298.19 | 887.82 | 124.36 |
| RealLight-max | 89.43 | 113.45 | 151.63 | 193.63 | 271.34 | 319.57 | 931.52 | 131.22 |
Values are means over ten evaluation runs with distinct random seeds.
Two caveats apply when reading this table. The Fixed-Time, SOTL, GRL, CoLight, PressLight, and IPDALight values are adopted from the IPDALight benchmark and were not rerun here, so implementation and stochastic-training differences cannot be ruled out. Dispersion statistics and significance tests are not reported, so the small differences in Jinan and Hangzhou should not be read as statistically significant.
All results are produced in simulation with exact vehicle counts. Sensing error, controller latency, and field operation are not evaluated.
The accompanying paper is under review at IEEE Access:
T. Eom, M. Park, S. Kim, and J. Yeo, "RealLight: Decentralized Networked Traffic Signal Control under Partial Observability."
Released under the MIT License. See LICENSE.