Challenges in Reward Design for Reinforcement Learning-based Traffic Signal Control - An Investigation Using CO2 Emission Objective

Watch a 21-minute conference talk exploring the challenges of designing effective reward systems for Deep Reinforcement Learning (DRL) in traffic signal control, with a focus on minimizing CO2 emissions at intersections. Dive into the complexities of training DRL agents using the SUMO (Simulation of Urban MObility) simulator, examining how different reward metrics and combinations affect agent performance. Learn why emission-based rewards prove inefficient for training Deep Q-Networks (DQN), discover the sensitivity of agent performance to reward parameter variations, and understand why certain reward formulations perform inconsistently across different scenarios. Explore key findings about reward properties that impact reinforcement learning-based traffic signal control, including the importance of informativeness and expressiveness in reward design. Follow along as presenters Christian Medeiros Adriano and Max Schumacher demonstrate practical examples like Cliff Walking, discuss the alignment of Traffic Signal Control agents with rewards, and share valuable insights about technologies that significantly enhanced their research outcomes.