Reproduce “CoordLight: Learning Decentralized Coordination for Network-Wide Traffic Signal Control” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yifeng Zhang, Harsh Goel, Peizhuo Li, Mehul Damani, Sandeep Chinchali, Guillaume Sartoretti
The CoordLight framework advances network-wide adaptive traffic signal control through a decentralized Multi-Agent Reinforcement Learning approach. It introduces Queue Dynamic State Encoding (QDSE) and a Neighbor-aware Policy Optimization (NAPO) algorithm, which together consistently reduce average travel time by up to 9.23% compared to leading baselines on real-world urban traffic datasets.
Key insights
- A predictive state representation like QDSE, which models queueing dynamics, enables agents to anticipate impending congestion and make proactive decisions, moving beyond purely reactive responses to current traffic conditions.
- Explicitly modeling and learning dynamic inter-agent dependencies through attention mechanisms within a decentralized MARL framework is crucial for achieving robust coordination and stable training in complex multi-intersection traffic control.
- Integrating neighbor-aware information into both the actor (policy) and critic (value) networks, via components like spatio-temporal aggregation and a state-action decoder, significantly enhances an agent's ability to learn cooperative strategies that optimize network-wide traffic flow without centralized control.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.