Cooperative Advisory Residual Policies for Congestion Mitigation

Hasan, Aamir; Chakraborty, Neeloy; Chen, Haonan; Cho, Jung-Hoon; Wu, Cathy; Driggs-Campbell, Katherine · 2024 · openalex

DOI: 10.1145/3699519

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of mitigating traffic congestion using cooperative advisory systems that provide real-time speed advice to human drivers. While fleets of autonomous vehicles (AVs) can effectively smooth traffic flow, their widespread deployment is hindered by safety risks, high costs, and the need for extensive centralized sensor infrastructure. Existing semi-autonomous solutions, such as Adaptive Cruise Control, often require complex infrastructure and fail to adapt to individual driver behaviors. The authors propose a class of learned "Cooperative Advisory Residual Policies" that advise a single human driver to adjust their speed to dampen stop-and-go waves. This approach aims to improve socioeconomic factors like commute time and fuel consumption while accounting for the uncertainty and diversity of human driver reactions to instructions. To develop these policies, the authors build upon Piecewise Constant (PC) policies, which provide constant speed advice for fixed time intervals to accommodate human reaction times. However, they identify three key drawbacks in PC policies: a mis-specified reward function that encourages tailgating, jerky action changes that are difficult for drivers to follow, and a lack of modeling for diverse driver behaviors. To address these, the authors introduce a novel reward function that incentivizes maximizing the speed of all vehicles while minimizing headway deviations, thereby promoting uniform traffic flow without tailgating. They also incorporate a term to reward small deviations between consecutive advised actions, making the advice smoother and easier to follow. Crucially, they introduce a driver policy model that simulates human instruction following by accounting for imperfect execution (noise), reaction time delays, and intentional speed offsets. This model allows the system to train in simulation with realistic human-like behavior. The residual policies are trained using reinforcement learning to offset the actions of the base PC policy, improving stability and training efficiency. The system further includes a Driver Trait Inference module using a variational autoencoder to learn latent representations of driver traits in an unsupervised manner, enabling the personalization of advice. The authors evaluate their approach through simulation experiments and a user study with 16 participants using a driving simulator. The simulation results demonstrate that the residual policies successfully mitigate congestion, achieving up to a 20% improvement in a combined metric of speed and speed deviations compared to baseline PC policies. The user study yielded even stronger results, showing a 40% improvement over baselines. Furthermore, the study confirmed that the policies are human-compatible and effectively personalize to individual driver traits, providing advice that drivers find easier to follow and more effective in reducing congestion. The significance of this work lies in its demonstration that learned advisory systems can effectively mitigate congestion using only a single vehicle with a human driver, without requiring fleet-level autonomy or extensive infrastructure. By explicitly modeling driver behavior and personalizing advice, the proposed residual policies bridge the gap between theoretical simulation performance and real-world applicability. The findings suggest that cooperative advisory autonomy can offer a practical, cost-effective solution for traffic management that adapts to the nuances of human driving, potentially improving traffic flow and reducing emissions in mixed traffic environments. This approach advances the field by validating learning-based advisory systems through empirical user studies, addressing a critical gap in prior research that lacked human-in-the-loop validation.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-10; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• traffic density

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Theoretical Contribution: computational model