Deep-Flow-Optimizer is an intelligent traffic management system built to solve complex urban congestion and bridge spillback. Utilizing Deep Q-Learning (DQN) and the SUMO simulator, this project demonstrates how AI agents can autonomously manage signal phases based on real-time lane-level data.
Urban intersections often suffer from downstream "spillback" that causes total network gridlock. This project benchmarks three control strategies:
- Fixed-Time: Traditional rigid signal cycles.
- Tabular Q-Learning: An RL agent using discrete state-mapping.
- Deep Q-Network (DQN): A neural-network-based agent for high-dimensional state spaces.
| Metric | Fixed-Time | Q-Learning | Deep Q-Network |
|---|---|---|---|
| Total Throughput | 154 veh | 157 veh | 165 veh |
| Network Delay | 381,959s | 283,253s | 472,157s |
| CO2 Impact | Baseline | -1.4% (Saving) | +5.3% (Learning Phase) |
- Install SUMO.
- Install Python dependencies:
pip install -r requirements.txt. - Run the DQL agent:
python DQL_Agent.py.
While the DQN agent maximized vehicle throughput (165 arrived), the Tabular Q-Learning agent achieved the highest overall efficiency with a 25.8% reduction in delay. This suggests that for single-intersection optimization, lower-dimensional state mappings converge more effectively within short training windows.