Modeling Outage Substitution for Robust Reinforcement Learning in Traffic Signal Control

Reinforcement learning for traffic signal control is evaluated almost exclusively on networks with reliable electrical power, an assumption that fails across much of West Africa. We study signal control under stochastic power outages on a five-junction traffic corridor reconstructed from OpenStreetMap geometry in Lagos, with demand calibrated to published Nigerian measurements. We train a deep Q-network that outperforms fixed-time, actuated and max-pressure control, then evaluate it under an outage process with three models of the behaviour that replaces the signal while it is dark. Outage cost depends on the substitution regime rather than on the outage itself. Under self-organised priority rules a dark corridor outperforms the trained controller by up to 57.9%, while under human direction delay rises by up to 648%. A sensitivity sweep over the human-direction model yields degradation between +157% and +1167% without changing sign, and the same asymmetry holds for non-learned controllers. We also find that every OpenStreetMap-signalised node in the study region controlled a single approach, which leaves adaptive control undefined without network reconstruction.