Cascaded Attention: Adaptive and Gated Graph Attention Network for Multiagent Reinforcement Learning

Learning Multi-Intersection Traffic Signal Control via Coevolutionary Multi-Agent Reinforcement Learning

“…This approach requires the construction of an interaction network that captures the attentional interactions of multiple signals. To model the interaction network, we employ a graph neural network [27], [33]- [35] where each signal is represented as a vertex, and the connections between them are represented as directed edges. Additionally, we propose that the relationships between all signals should co-evolve over time in light of our novel insights into the temporally stable traffic pattern.…”