acceptodds
Under review as a conference paper at ICLR 2027

CAPE: Closed-Loop Adaptive Planning and Execution for Multi-Node Traffic Fault Recovery

Abstract

Traffic fault recovery aims to restore network performance after a fault. Existing approaches follow a diagnose-then-control pipeline: localize a root cause and control there. This overlooks two properties: the root cause need not be the most effective control location, and a single intervention cannot restore a network whose congestion keeps evolving. We introduce CAPE (Closed Loop Adaptive Planning and Execution), a closed-loop framework that adjusts interventions as traffic evolves. CAPE uses two-stage decision making: a Planner selects control nodes and measures, and an Executor generates safe actions, enabling effective search in a large hybrid action space. For rare control sets, CAPE calibrates a low-dimensional strategy code via paired simulations and distills it into a generator, improving execution without fine-tuning the full Executor. Across five networks and four fault types, CAPE achieves an average Recovery Gain of 89.70%, exceeding the strongest baseline by 9.15 percentage points. Code is available at https://anonymous.4open.science/r/CPAE-2785.

Primary area:
reinforcement learning
Keywords:
traffic fault recovery, traffic signal control, closed-loop planning, multi-node control, reinforcement learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Ah4ROpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper focus on a real-world problem in traffic signal control. When one node is controlled by an intelligent agent, the traffic will go to the downstream node and cause traffic jam. The author propose a novel method to solve this problem, the method propose an adaptively policy to add control node and adjust traffic to improve the recovery performance after node fault. The experiments are comprehensive and presentation is good.

Critical strengths
  1. This paper focus on a real-world traffic problem which is very important to research. The author propose CAPE to solve this problem by adaptively adding control node and adjust traffic to improve the recovery performance after node fault.

  2. The experiments are comprehensive, the author design a fault-recovery experiment to test the recovery ability for different method, which is convincing.

  3. The presentation looks good and enough to publish in ICLR.

Critical weaknesses
  1. Images should enhance the visual presentation. For example, we should use the same colour for one module and different colour for different moudle, of course, this is just my suggestion for your consideration.

  2. The counterfacual rollouts may be difficult to obtain in real-world traffic environments. This raises concerns about the practicality and general applicability of the proposed online adaptation mechanism.

Questions for authors
  1. Can the author have a more deep think about the counterfactual training problem? Because in the real-world this is difficult to simulate. I found a paper called CFLight[1] which uses counterfactual learning to improve the safety, the author can refer it.

  2. Can the author investigate more papers for this problem and compare with your method? I found some papers called MalLight, DiffLight, FuzzyLight[2,3,4] which has the similar problem with yours. I would like to improve my score if the author can resolve these problems.

[1] CFLight: Enhancing Safety with Traffic Signal Control through Counterfactual Learning. KDD, 2026.

[2] MalLight: Influence-Aware Coordinated Traffic Signal Control for Traffic Signal Malfunctions. CIKM, 2024.

[3] Difflight: a partial rewards conditioned diffusion model for traffic signal control with missing data. NeurIPS 2025.

[4] FuzzyLight: A Robust Two-Stage Fuzzy Approach for Traffic Signal Control Works in Real Cities. KDD 2026.

Reviewer_nzx8OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper presents CAPE, a closed-loop framework for multi-node traffic fault recovery. A Planner dynamically selects control locations and measures, while an Executor generates continuous actions. Disagreement between their value estimates triggers strategy-code calibration through paired simulations, followed by distillation for reuse. Experiments across five SUMO networks and four fault types evaluate recovery performance, efficiency, and robustness.

Critical strengths

[S1] The paper addresses a practically relevant challenge: intervention locations and measures need to adapt as congestion propagates. The Planner–Executor decomposition provides a coherent way to separate control-set selection from continuous action generation. [S2] Experiments across five networks and four fault types report consistent recovery improvements. Multi-seed evaluations, component ablations, calibration-schedule comparisons, and robustness analyses provide useful evidence supporting the proposed design.

Critical weaknesses

[W1] The technical positioning and core decision procedure remain insufficiently specified. The introduction provides limited differentiation from the adaptive and hierarchical approaches discussed in Section 2. More importantly, the paper does not clearly specify how period-level rewards are assigned to internal set-edit transitions, or how the recomputed Executor advantage affects subsequent Planner decisions. These missing links make the central planning–calibration mechanism difficult to assess.

[W2] The recovery–efficiency trade-off is not sufficiently substantiated. Appendix B.2 describes paired simulations through the remaining recovery horizon. Although Appendix C.1 defines observation-to-action latency, the paper does not report calibration-triggered decision latency or specify the simulation hardware and parallelism underlying the aggregate results. Comparable online simulator access and intervention budgets across baselines also remain unclear, limiting assessment of the reported gains under comparable resource conditions.

[W3] In the inspected public implementation, the training loop does not invoke the defined Planner/Executor SAC updates, the standalone evaluation script constructs a new agent without loading a trained checkpoint, and the rollout sets both normal and fault reference delays to the same first-step value, inconsistent with Eq. (26). These discrepancies require reconciliation with the exact version used to produce the reported results.

[W4] There are too many hyperparameters that are set to fixed values without justification. For instance, in Eqs. 2, 13, and 16, specific hyperparameters such as c_t = 0.1 are used, but it is difficult to understand where these numbers come from. The authors should provide an ablation study or theoretical motivation for these choices, or at least discuss their sensitivity.

Questions for authors

[Q1] How are period-level rewards assigned to the Planner’s internal set-edit transitions, including Stop? How does the recomputed Executor advantage enter the Planner’s subsequent decision rule?

[Q2] Under the observation-to-action latency definition in Appendix C.1, how are paired simulations through the remaining recovery horizon, local calibration, and distillation included in the reported timing? Please provide latency for calibration-triggered decisions and specify the hardware and parallelism. Please also clarify the baselines’ online simulator access, available control measures, and intervention budgets.

[Q3] Which code version and checkpoints produced the reported results? Please reconcile the discrepancies concerning SAC updates, checkpoint loading, and Recovery Gain reference initialization in the inspected public release with the actual experimental pipeline, and clarify whether any reported results require recomputation.

Reviewer_tmY7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CAPE, a closed-loop framework for multi-node traffic fault recovery. It employs hierarchical RL to dynamically select control nodes and measures, while an Executor generates continuous actions for signal timing, flow control, and path diversion. An advantage-disagreement calibration mechanism further improves execution for rarely visited control sets.

Critical strengths
  1. The paper explores a relatively new task of dynamic multi-node traffic fault recovery, extending conventional traffic control to more complex fault scenarios.
  2. The experiments are comprehensive, covering five traffic networks and four fault types.
Critical weaknesses
  1. CAPE combines hierarchical RL, dynamic control-set selection, and online strategy calibration. However, hierarchical RL and structured action generation have been extensively studied in prior work. For example, AlphaStar employs autoregressive selection of variable-sized unit sets for complex action generation. The authors should better distinguish the methodological contributions of CAPE from existing approaches, particularly regarding its hierarchical Planner–Executor design and dynamic control-set selection.
  2. The relationship between traffic fault recovery and conventional multi-intersection TSC is not sufficiently explained. The proposed framework appears to extend multi-intersection TSC by incorporating additional control actions, such as flow control and path diversion. The authors should better clarify its distinction from existing TSC approaches, including studies that jointly optimize traffic signals and vehicle-level control. Moreover, the practical feasibility of implementing these additional control measures in real-world traffic networks should be discussed.
  3. CAPE jointly optimizes signal timing, flow control, and path diversion, whereas most baselines were originally designed for traffic signal control. This raises concerns about the fairness of the comparisons.
Questions for authors
  1. What are the key methodological advantages of CAPE over existing hierarchical RL and structured action generation approaches?
  2. How does CAPE differ from existing multi-intersection TSC and joint signal–vehicle control methods? How feasible are the proposed control measures in real-world traffic networks?
  3. Do all baselines have access to the same control capabilities as CAPE, including signal timing, flow control, and path diversion? How is a fair comparison ensured?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.