SSTLight: Selective State-Space Trajectory Learning for Offline-to-Online Traffic Signal Control

Offline-to-online learning reduces the reliance of traffic signal control on extensive online interaction, but return-conditioned traffic trajectories require effective temporal representations of evolving traffic states. We propose SSTLight (Selective State-Space Trajectory Learning for Traffic Signal Control), which organizes return-to-go, local observations, and historical signal actions into a causally interleaved return–observation–action trajectory and encodes it with an input-dependent Selective State-Space Model (Selective SSM). Selective parameters regulate latent-state updates and readout according to current trajectory content, producing a contextual representation for signal-action prediction. SSTLight combines offline teacher training, teacher–student knowledge distillation, and online student adaptation under decentralized independent control. Across eight public Simulation of Urban MObility (SUMO) scenarios, spanning synthetic and real-road-network settings, with 1–16 controlled intersections and five experimental runs, SSTLight achieves a lower mean Average Delay than DTLight in seven scenarios, with a macro-average relative reduction of 4.10%. In a controlled encoder comparison with matched data, trajectory interfaces, training budgets, and online update scope, Selective SSM achieves a lower mean Online Best in all eight scenarios, with reductions of 0.64–18.98% and a macro-average of 4.26%. Approximate parameter matching on Cologne1, Cologne3, and Cologne8 preserves the same favorable mean-performance trend. These results provide empirical evidence that Selective SSM trajectory encoding can serve as a temporal encoding alternative to causal self-attention for return-conditioned offline-to-online traffic signal control.