#352 Evaluating Resilience in Mixed-Autonomy Transportation Systems

Principal Investigator
Carlee Joe-Wong
Start Date
Jan. 1, 2021
End Date
June 30, 2022
Project Type
Research Applied
Grant Program
FAST Act - Mobility National (2016 - 2022)
Grant Cycle
2021 Mobility UTC "Big Idea"


Future deployments of autonomous vehicles raise questions on how the actions of such vehicles may affect transportation systems as a whole, including the human-driven vehicles with which they share the road. We propose to build a model of how autonomous vehicles can affect such mixed-autonomy systems and in particular their resilience to traffic disruptions. We will validate our results with simulations on Pittsburgh data and derive policy recommendations for autonomous vehicle regulations.    
As autonomous vehicles near widespread deployments, we can envision a future where an increasing fraction of vehicles on the road will have some degree of autonomy. While many works have studied how autonomous vehicles might coordinate with each other to enhance transportation safety and efficiency, relatively few studies have examined interactions between human-driven and autonomous vehicles. One might expect that the safety and efficiency benefits of autonomous vehicles increase in proportion to their prevalence on roads, but the magnitude of this effect may be more complex; e.g., a threshold fraction of vehicles may need to be autonomous to have a measurable effect on safety or efficiency. Quantifying these benefits may aid regulators and policymakers in estimating the true benefits that autonomous vehicles can bring to transportation networks, particularly when the prevalence of autonomous vehicles can vary throughout a city’s road network.

This work may further illuminate the effects of autonomous vehicles on providing transportation in disadvantaged communities. It is likely, for example, that autonomous vehicles will initially be more prevalent in wealthier neighborhoods. Our work will examine how this imbalance translates into different traffic patterns across the transportation network. Autonomous vehicles can moreover provide mobility to people with disabilities who cannot drive conventional vehicles, and our work may facilitate their adoption by developing policies for autonomous vehicles to follow in choosing which routes to take, etc. so as to benefit the transportation network as a whole.

In this project, we will consider the particular scenario of cascading failures in road networks: when a road or service fails (e.g., due to construction, a subway outage, etc.), traffic is re-routed to other roads or services, which may in turn be overwhelmed, leading to a cascade of failures. Autonomous vehicles may be able to redirect themselves so as to stop or limit these cascades. In particular, compared to human-driven vehicles, autonomous vehicles may be able to access more up-to-date information, thus allowing them to take more informed actions in response to road failures, e.g., avoiding alternate routes that are already becoming congested. Autonomous vehicles may also allow for more coordination than human-driven ones; e.g., if a company deploys a fleet of autonomous vehicles, a remote coordination system can direct them to take different routes around a crowded road. Our goal is to quantify the impact of cascading failures in terms of the fraction of autonomous vehicles present, paying particular attention to cases where the autonomous vehicles are concentrated in specific areas of the network, e.g., more affluent areas of a city, and the resulting citywide effects.

We plan to use a cascading failure model based on flow redistribution over a network. Each edge of the network, representing a road segment, is associated with a delay function that depends on its current flow (i.e., number of cars currently using the road). There is also a capacity assigned to each edge indicating an upper limit on the flow it can sustain. If a road segment fails (due to an accident or closure), the flow demand would be redistributed to other roads; i.e., cars are re-routed to alternative roads. Upon this redistribution, the capacity on otherwise lightly-loaded edges may be exceeded leading to further failures and further redistribution of flows to other routes, only to face a similar consequence. This project will investigate how such cascading failures can be avoided with the help of autonomous cars (by leveraging their capabilities for coordinated redistribution actions).

We plan to experiment our models/algorithms and validate our results in Pittsburgh regional networks. The multi-class network model and its data-driven framework will be developed based on Mobility Data Analytics Center - Prediction, Optimization, and Simulation Toolkit for transportation Systems (MAC-POSTS) previously supported by Mobility21 projects. MAC-POSTS is not only a mesoscopic traffic simulation software in the road network, but also a passenger/vehicle behavioral modeling package in the general roadway network. MAC-POSTS is capable of modeling a comprehensive real-world mobility network with mixed (cars versus trucks, autonomous vehicles versus human driven vehicles) traffic flow, multi-class network, heterogeneous travelers route choices. The mobility model can be calibrated with multi-source datasets efficiently. Those datasets include, but are not limited to, traffic counts, traffic speed, incidents, and weather conditions, which have been archived by the Mobility Data Analytics Center (MAC) led by co-PI Qian. The team has established a regional dynamic network model for the Pittsburgh Metropolitan Area that simulates all vehicle trips from their respective origins to destinations in high spatio-temporal resolutions. We will then leverage the existing modeling efforts to categorize autonomous vehicles and human-driven vehicles by their respective driving behavior and routing strategies, and to simulate how resilient each community is to cascading disruptions under different traffic flow propagation patterns. Since traffic may vary by day from the average counts, we will use randomized perturbations from the average data to assess the robustness of our algorithms to different initial load distributions.

We finally note that autonomous vehicles may be able to benefit transportation systems in more ways than just preventing cascading failures. For example, autonomous vehicles require less (or no) attention from drivers and thus may be able to take longer routes than human-driven ones without inconveniencing their passengers, thus reducing the overall road congestion. While some initial works have explored this routing problem, they have not fully considered the separate effects of factors like autonomous vehicles’ access to up-to-date congestion information and an ability to centrally control autonomous vehicles’ movements. Autonomous vehicles may also have more flexibility in delaying their departure times and choosing parking spaces far from their destinations, which may improve road and parking congestion. We plan to explore such ideas in future work, building on the simulator and mobility models that we will develop in this proposal.
We envision the project to last 18 months. We break down the proposed research into three six-month long stages:

Months 1 - 6: Model formulation and analysis.
We will begin the project by formulating the cascading failures model when autonomous vehicles are present. We will then derive bounds on the robustness of the network to different degrees of disruption, when the human-driven and autonomous vehicles follow different load propagation patterns.

Months 7 - 12: Simulations on traffic data traces.
The next step of the project will be to simulate our models and algorithms using real traffic counts, as detailed in the project description. Our goal will be to first validate our theoretical performance bounds under different degrees of disruption, and then evaluate which load propagation patterns show the most resilience, given different penetration rates of autonomous vehicles.

Months 13 - 18: Model refinement and policy implications.
In the last six months of the project, we will refine our models to better match our findings in simulation, and begin extending the models to study how autonomous vehicles in one part of the network can have larger effects on the rest of the transportation network. In partnership with the Southwestern Pennsylvania Commission (see attached collaboration letter), we will also examine the policy implications of our results, e.g., whether they specify how autonomous vehicles should react to traffic disruptions so as to avoid cascading failures.
Strategic Description / RD&T

Deployment Plan
Our first deployment product will be an open-source simulator that captures cascading traffic effects due to disruptions. Specifically, it allows us to simulate individual vehicles (including cars, trucks, human-driven, autonomous, etc.) entering the road network at a specified time, and choosing a route between a given origin and destination. We can thus simulate the re-routing actions of each vehicle upon encountering traffic congestion or road blockages, and the resulting effects on congestion for nearby roads. We can record the travel time of each simulated vehicle on each road traveled. As detailed in the project description, the simulator will be based on the existing MAC-POSTS simulator developed by co-PI Qian. Our simulator will be integrated with Open Street Map and Pittsburgh traffic data, which may be replaced with other data sources by other researchers. This simulator can also be used to study the effects of autonomous vehicles in scenarios beyond cascading failures, including reductions in congestion due to autonomous vehicles choosing more efficient routes between their origin and destination.

We also expect that our results will have implications for regulators aiming to reduce congestion on Pennsylvania roadways. Our second deployment effort will be to share our findings with Pennsylvania transportation agencies (see attached support letter from the Southwestern Pennsylvania Commission) so that they may impact the regulatory policies developed for autonomous vehicles in future deployments. For example, autonomous vehicle fleets might be required to coordinate with each other to reroute themselves around blockages of certain roads, if doing so can prevent cascading failures. We plan to spend part of the last six months of the project on examining the policy implications of our findings.

In addition to these products, we also expect to publish our findings in appropriate conference or journal venues (e.g., IEEE Transactions on Networking, IEEE Transactions on Intelligent Transportation Systems). Preprints of these publications will be publicly released online.
Expected Outcomes/Impacts
Our expected accomplishments are (1) a model for cascading failures in transportation systems, (2) an analysis of how these failure patterns change as a function of the autonomous vehicles present and how they re-route themselves when encountering a blocked or congested road, (3) validation of these analytical results on a traffic simulator that we plan to build, and (4) an enumeration of policy recommendations and implications based on our analytical and simulation findings.

Our primary metric will be the amount of disruption that the network can withstand when autonomous vehicles are present, i.e., the number of roads that are not overwhelmed by traffic exceeding the road capacity. We will also consider metrics such as the average travel time experienced by human-driven and autonomous vehicles in the presence of such disruptions.

The second category of metrics will be the accuracy of our analytical results in simulation, which can measure how useful our analytical findings might be in practice. We will define these metrics depending on the analytical results that we are able to obtain.
Expected Outputs



Individuals Involved

Email Name Affiliation Role Position
eelumar@andrew.cmu.edu Elumar, Eray Can Carnegie Mellon University Other Student - PhD
cjoewong@andrew.cmu.edu Joe-Wong, Carlee Carnegie Mellon School of Engineering PI Faculty - Untenured, Tenure Track
ichengl@andrew.cmu.edu Lin, I-Cheng Carnegie Mellon University Other Student - PhD
seanqian@cmu.edu Qian, Sean Carnegie Mellon University Co-PI Faculty - Untenured, Tenure Track
oyagan@ece.cmu.edu Yagan, Osman Carnegie Mellon University Co-PI Other


Amount of UTC Funds Awarded
Total Project Budget (from all funding sources)


Type Name Uploaded
Data Management Plan DataManagementPlan_dXVAKth.pdf Nov. 25, 2020, 9:01 a.m.
Progress Report 352_Progress_Report_2021-03-31 March 31, 2021, 12:52 p.m.
Progress Report 352_Progress_Report_2021-09-30 Sept. 30, 2021, 9:47 a.m.
Publication ASC: Actuation system for city-wide crowdsensing with ride-sharing vehicular platform Feb. 9, 2022, 5:52 a.m.
Publication ilocus: Incentivizing vehicle mobility to optimize sensing distribution in crowd sensing Feb. 9, 2022, 5:54 a.m.
Publication Dynamic Coupling Strategy for Interdependent Network Systems Against Cascading Failures March 29, 2022, 5:03 a.m.
Progress Report 352_Progress_Report_2022-03-30 March 29, 2022, 5:03 a.m.
Final Report 352_Final_Report.pdf July 25, 2022, 5:51 a.m.
Publication Mixed-Autonomy Era of Transportation: Resilience & Autonomous Fleet Management Sept. 30, 2022, 7:30 p.m.
Presentation Enhancing Robustness for Interdependent Network Systems with Dynamic Coupling Strategy Sept. 30, 2022, 7:30 p.m.
Progress Report 352_Progress_Report_2022-09-30 Sept. 30, 2022, 7:30 p.m.
Publication Gcn-se: Attention as explainability for node classification in dynamic graphs March 30, 2023, 5:44 a.m.
Publication A neural-based bandit approach to mobile crowdsourcing March 30, 2023, 5:45 a.m.
Publication Modems: Optimizing edge computing migrations for user mobility March 30, 2023, 5:45 a.m.
Publication Plan B: Design Methodology for Cyber-Physical Systems Robust to Timing Failures March 30, 2023, 5:46 a.m.
Publication On merits and viability of multi-cloud serverless March 30, 2023, 5:51 a.m.
Publication Taxi-for-all: Incentivized Taxi Actuation System for Balanced Area-wide Service March 30, 2023, 5:52 a.m.
Publication Incentivizing Opportunistic Data Collection for Time-Sensitive IoT Applications March 30, 2023, 5:53 a.m.
Publication Interpretable clustering on dynamic graphs with recurrent graph neural networks March 30, 2023, 5:53 a.m.
Publication On the connectivity and giant component size of random k-out graphs under randomly deleted nodes March 30, 2023, 6:15 a.m.
Publication Evaluating Resilience in Mixed-Autonomy Transportation Systems March 30, 2023, 6:21 a.m.
Publication Evaluating Resilience in Mixed-Autonomy Transportation Systems [supporting dataset] March 30, 2023, 6:22 a.m.

Match Sources

No match sources!


Name Type
Southwestern Pennsylvania Commission Deployment Partner Deployment Partner
Drive Sally Deployment Partner Deployment Partner