Conditional Effects of Connected-Vehicle Requests in Reinforcement-Learning Traffic Signal Control

Abstract

Connected vehicles can provide upstream observations and priority cues, but a request mechanism is useful only if it improves decisions beyond queue-responsive adaptation. We test this distinction at one site-derived SUMO intersection with a request-aware deep Q-network (DQN) and a matched queue-only ablation. The variants share the architecture, base congestion objective, training schedule, and 10–20~s timing shield; the request-aware variant additionally incorporates request-state inputs and an auxiliary request-shaped reward term. Five training realizations per variant were evaluated against fixed-time, presence-actuated, and max-pressure control at three demand levels, seven connected-vehicle penetrations, and ten paired traffic realizations, yielding 2,730 one-hour runs.

To represent both stochastic sources, we independently resample the training and traffic axes of each $5\times10$ crossed result matrix. Queue-only DQN reduced mean queue relative to each of the three conventional baselines in all 21 demand–penetration conditions, with all 21 crossed-bootstrap intervals per baseline (63 in total) excluding zero. The request-aware DQN had lower point estimates than the same baselines in all conditions, but one interval at nominal demand and 50% penetration included zero because training-realization dispersion was large. Relative to queue-only control, the 95% crossed intervals excluded zero in only eight of 21 conditions: three favored the complete request mechanism (Full DQN) and five favored Queue-only DQN. At 0% penetration, no request occurs online, so that contrast is a training-history diagnostic rather than an online request effect. Request counts show that policies approved 75–81% of hold requests and 27–37% of switch requests; these diagnostics describe policy behavior but do not identify the cause of performance differences. Across this sweep, queue-responsive adaptation is the reproducible gain; the incremental effect of request conditioning is demand- and training-sensitive and is not monotonic in penetration.

🚀 Key Empirical Insights

1. Adaptivity Advantage: Both Full and Queue-only DQN substantially outperform Fixed-time, Actuated, and Max-pressure baselines across all 21 demand $\times$ penetration conditions (queue reduction of $0.40$–$0.93$ veh/lane; all 63 crossed-bootstrap 95% CIs exclude zero).
2. Request Fragility: Adding the complete request mechanism does not monotonically improve performance. At nominal design demand (1,000 veh/h), Full DQN exhibits longer queues than Queue-only DQN across 0%–80% penetrations.
3. Demand Inversion: The marginal effect of request conditioning flips sign with demand level: beneficial under low demand ($700$ veh/h at 20% and 80%), but detrimental or neutral under nominal and saturated regimes.
4. Approval Asymmetry: Trained policies approve 75%–81% of HOLD requests but only 27%–37% of SWITCH requests, causing request-induced state perturbations that conflict with queue clearing.
5. Retraining Dispersion: Training stochasticity dominates traffic stochasticity in specific regions (e.g. 50% penetration), causing wide confidence bands due to outlier policy convergence.

🔗 Links & Resources

• GitHub Repository: AnXu-ITS/cv-request-dqn-tsc
• Preprint Manuscript (PDF): Download Full Paper (13 Pages)

📖 BibTeX Citation

@article{xu2026conditional,
  title={Conditional Effects of Connected-Vehicle Requests in Reinforcement-Learning Traffic Signal Control},
  author={Xu, An},
  journal={Preprint},
  year={2026},
  url={https://github.com/AnXu-ITS/cv-request-dqn-tsc}
}

Recommended citation: Xu, An. (2026). "Conditional Effects of Connected-Vehicle Requests in Reinforcement-Learning Traffic Signal Control." Preprint.
Download Paper