This project explores the use of Deep Reinforcement Learning (DRL) for intelligent traffic light control at a single four-way intersection, simulated using SUMO (Simulation of Urban Mobility).
Our goal was to minimize traffic congestion and vehicle waiting time by training an agent that learns to dynamically adjust traffic signal phases based on real-time conditions.
We designed and evaluated multiple RL-based traffic control strategies:
- Agent selects a traffic phase; duration is fixed (e.g., 30s).
- Techniques: DQN, Double DQN, Dueling DQN, PER
- Agent selects both phase and duration from a discrete set.
- Improves adaptability during varying traffic loads.
- Agent switches phases based on vehicle throughput (e.g., % of cars that have passed).
- Yields the best performance in terms of reduced wait times and queue lengths.
- 🧠 DQN, Double DQN, Dueling DQN
- 🧪 Prioritized Experience Replay (PER)
- 🎯 Curriculum Learning
- 📈 Layer Normalization, Dropout, Gradient Clipping
- ⚙️ Dynamic Action Spaces (phase + duration or throughput thresholds)
- 🧮 PyTorch for neural network implementation
- 🛣️ SUMO for traffic simulation
- 🧪 Python for RL environment integration
Our best-performing agent (throughput-based control) achieved:
- 🚗 ~60% reduction in waiting time
- 🚦 Significant improvement in traffic flow and queue management
- 💡 Adaptive behavior across varied traffic patterns