Project: #352 Evaluating Resilience in Mixed-Autonomy Transportation Systems Progress Report - Reporting Period Ending: Sept. 30, 2021 Principal Investigator: Carlee Joe-Wong Status: Active Start Date: Jan. 1, 2021 End Date: June 30, 2022 Research Type: Applied Grant Type: Research Grant Program: FAST Act - Mobility National (2016 - 2022) Grant Cycle: 2021 Mobility UTC "Big Idea" Progress Report (Last Updated: Sept. 30, 2021, 9:47 a.m.) % Project Completed to Date: 50 % Grant Award Expended: 0 % Match Expended & Document: 0 USDOT Requirements Accomplishments The objective of this project is to evaluate how autonomous vehicles can improve the resilience of transportation networks to exogenous disturbances like road construction, traffic accidents, and congestion. We plan to use our analysis to develop guidelines for regulators to specify the behavior of autonomous vehicles in a road network, so as to maximize resilience. To accomplish this goal, we plan to complete three major tasks. First, we will formulate a model of cascading failures in road networks, which captures the effects of users redistributing themselves around a disturbance, which can then cause congestion on nearby roads, leading to a cascade of road congestion and potentially failure of the network overall. This model will include the behavior of autonomous and human-driven vehicles; they may react differently to disturbances in transportation networks. Second, we will validate our models with simulations on road traffic simulators that can capture the behavior of autonomous as well as human-driven vehicles. Finally, we will translate our findings into policies that regulators may be able to impose on autonomous vehicles as they are more widely deployed. Over the past six months, we have made progress on our first task of formulating the cascading failure model for transportation networks. As summarized in our last report, we have considered the case of two networks that operate in parallel (for example, a road network and train network), where work can be offloaded or transferred between the two. Our work proposes a dynamic strategy in which the proportion of workload moving between the two networks may change depending on the current state of the network. We have extended our previous work on this strategy to evaluate its performance on networks that are not fully connected (specifically, Erdos-Renyi and Barabasi-Albert topologies), which are more similar to the topologies of transport networks. We show that on such networks, our proposed dynamic strategy still outperforms static transfer strategies in the sense that the network can survive larger initial failures. The magnitude of improvement relative to static strategies, however, is larger for topologies that are more similar to fully connected networks. We have further extended our two-network model to consider the resilience of a single network to initial failures, e.g., if roads are blocked, nearby roads may also experience congestion. Our goal in doing so is to model the different redistribution actions for autonomous and human-driven vehicles. The key difference in their responses is that autonomous vehicles (i) have more information about the current condition of all road segments, and (ii) are better able to coordinate with each other so as to optimize the behavior of the entire system, instead of their individual experience. Thus, we used a Stackelberg game formulation in which the AVs are the game “leaders,” making decisions on how to distribute themselves around congestion or blockages so as to maximize the collective system performance (e.g., minimizing overall travel time). These decisions account for the reactions of human users, who are modeled as the “followers” and make myopic, selfish re-distribution decisions. We have simulated this initial setting in a grid network and found that AVs’ intelligent decisions can alleviate failures. Moreover, even a limited number of AVs are able to make a significant difference in network failure, e.g., the number of surviving road segments (that are not blocked by congestion) is similar for 50% or 100% AVs, even though the former scenario includes many human-driven vehicles making selfish decisions. Our goal for the next six month reporting period is to continue to develop our cascading failure model, and to begin aligning our results with policy improvements. We had an initial discussion with Chuck Imbrogno from the Southwestern Pennsylvania Commission on how the preliminary results from our work might be applied to public policies. As a result of this discussion, we plan to focus on quantifying the effects of AV decisions for different prevalence rates, and on the policies that might be implemented to incentivize AV fleets to adjust their routing so as to reduce overall network congestion instead of only their travel times. Impacts Our work on analyzing how users should transfer from one transportation network to another so as to avoid cascading failures can help network operators identify ways in which they can avert failures by diverting users to an alternative mode of transport. We plan to submit these results to the IEEE Transactions on Network Science and Engineering. We expect our Stackelberg game formulation and analysis to yield insights into how regulators can induce AV fleets to act so as to reduce overall network congestion. In our preliminary meeting with Chuck Imbrogno, we received feedback that vehicle fleets are likely a small fraction of overall traffic in southwestern Pennsylvania in general, though Uber or Amazon may have significant impact in local regions like downtown Pittsburgh. We held a similar meeting to obtain feedback from representatives of the City of Pittsburgh, etc. on September 27. The main takeaway was that the assumption of AVs acting as a fleet to improve social welfare may not be realistic. AVs may have different trip purposes and may require different types of incentives (e.g., toll discounts or adjusted speed limits) to act in such a way so as to improve system resilience and respond to disruptions. We have sent out a summary slide deck to attendees and other stakeholders who could not attend the meeting, in order to receive more feedback from a larger range of regulatory agencies at the city, county, and state levels. Other No other outputs for this reporting period. Outcomes New Partners None to report. Issues No changes to report.