A Comparison of Reinforcement Learning Agents Applied to Traffic Signal Optimization
summary
A small OpenAI Gym wrapper called SUMO gym uses TraCI to train DQN and PPO agents on a single intersection in SUMO for 200 episodes (each 100,800 seconds); training reaches near-optimal policies in minutes on AWS. Both RL agents beat Webster and SUMO actuated baselines: no car waits >50s and queues rarely exceed ~100 m, but PPO overfits and becomes unstable on the dynamic P4 pattern while DQN generalizes better and converges to a different yet comparably effective policy.
AI-written summary · about
chapters (16)
- 0:00Introduction
- 0:42Why Reinforcement Learning
- 1:21Objectives
- 1:40Intersection
- 2:32Traffic Patterns
- 3:13Benchmark
- 3:58Framework Overview
- 5:46Demonstration
- 6:24Results
- 7:21Wall Time
- 7:38Waiting Times
- 8:28Speed
- 9:00Q lengths
- 9:25Performance
- 11:38State Transition matrices
- 14:10Conclusion
related talks
21mChallenges in Reward Design for Reinforcement Learning-based Traffic Signal Control
11mUsing Deep Reinforcement Learning to Coordinate Multi-Modal Journey Planning with Limited ...
20mEnergy Optimized Green Light Assist in Varying Traffic Scenarios Using Reinforcement Learning
35mCrowd dynamics, RL and Unity: A Journey
01h 17mTraining intelligent game agents using deep reinforcement learning
18mMaximizing Luck in Reinforcement Learning
18mSUMO Simulations for Federated Learning in Communicating Autonomous Vehicles
29mReinforcement Learning with Human Feedback