UC Berkeley

The digitization and automation of transportation systems is fundamentally transforming the transportation landscape, creating important opportunities and challenges for improv- ing the sustainability of transportation worldwide. There are more travel options to choose from than ever before, with real-time information on the cost and time trade-offs between alternative routes and modes of travel available in the palm of a hand. In particular, shared on-demand mobility services such as transportation network companies (TNC) (e.g., Lyft, Uber), bikesharing, and microtransit offer affordable and convenient alternatives to personal auto ownership that can complement or supplement existing public transit services. In addition, these services may aid in reducing existing inequities in access to fast and reliable transportation. However, despite the potential that innovative shared mobility service mod- els bring forth to improve the sustainability of transportation, the inefficiencies of fleet-based services such as TNCs and the low adoption rates of pooled rides that transport multiple travelers in the same vehicle have contributed to worsening congestion in several regions across the United States. Meanwhile, the design and deployment of transportation demand management (TDM) strategies has not kept pace of disruption nor the corresponding evolution in travel behavior.

In light of the pressing need to improve the sustainability of a rapidly evolving transportation ecosystem, this dissertation contributes to the theory, methodology, and state of knowledge of optimal mechanism design for multi-faceted TDM strategies. With a focus on congestion pricing strategies, this research aims to facilitate the design and analysis of data-driven TDM strategies that incorporate a multitude of policy levers (e.g., congestion pricing, multi-modal incentives, public transit operations) using a simulation-based multi-objective optimization approach to inform decision-makers about the inherent trade-offs between congestion and emission reductions, economic feasibility, and transportation equity.

In particular, I focus on the optimization of congestion pricing and targeted incentive schemes in multi-modal transportation networks including pooled ride options. Congestion pricing aims to reduce congestion by charging road users for driving on congested roads. A review of the various aspects of the transportation pricing optimization problem as studied in the literature is presented in Chapter 2, including the specification and analysis of various dimensions of demand sensitivity, congestion pricing structure and charging zone design, optimization objectives, optimization approaches, and transportation equity analysis. Studies of the optimization of pricing structures and charging zones for congestion pricing schemes have established that greater toll levels and charging zone coverage produce greater reductions in driving, which is deemed beneficial with respect to total system-wide travel time reductions. However, as is pointed out by public acceptance studies and literature on the equity implications of congestion pricing, the cost burden of congestion mitigation is disparately borne by lower-income individuals who are the most likely to be financially incentivized to adopt less desirable alternatives to driving. Few congestion pricing optimization studies have incorporated transportation equity objectives; none have included equity in addition to other efficiency and environmental objectives. The contributions of this dissertation to the literature on congestion pricing optimization span the theory of optimal mechanism design for multi-objective congestion pricing strategies, methodology for simulation-based multi-objective optimization of multi-faceted TDM strategies, and empirical understanding of congestion pricing strategies optimized with respect to multiple policy objectives.

Mechanism Design for Optimal Congestion Pricing PoliciesChapter 3 establishes that equity-based objectives and the inclusion of monetary incentives for the adoption of driving alternatives are feasible strategies for tackling the equity issues inherent in congestion pricing optimization (i.e., the disparate distributions of increased costs for lower income drivers and reduced travel times for higher income drivers). In this chapter, I formulate a bi-level optimization problem to compute optimal link- and mode-specific tolls and targeted mode-specific incentives (i.e., direct monetary transfers) using aggregate-level information on the flows of network users using various modes of transportation. This work contributes to the theory of congestion pricing optimization by proving the existence of optimal multi-modal congestion pricing schemes including both tolls and monetary incentives that are optimized with respect to equity-focused objectives defined on the basis of the distributional impacts of the pricing scheme across travelers. Several functions inspired by different theories of justice are presented as alternative social objective functions, including: 1. Utilitarian: maximize the sum of individual utility (i.e., a quasilinear function of travel time, cost, and other factors weighted by the individual demand sensitivity to each) of travel, 2. Egalitarian: maximize the sum of individual utility using the average demand sensitivity to weight travel time cost, 3. Equality of Opportunity: maximize a weighted sum of the individual utility, again using the average demand sensitivity, with weights applied to the overall utility that are scaled according to a societal ordering of disadvantage across network users (e.g., weights decrease with income, weights increase for residents of historically underserved neighborhoods or members of historically underserved communities), and 4. Rawlsian: maximize the minimum individual utility of travel, again using the average demand sensitivity. Consideration of the three latter social choice functions offers a deviation from the traditionally utilitarian view on congestion pricing optimization. Implementation of the bi-level optimization problem presented in chapter 3 and analysis of the impacts of each of the alter- native objective functions on the optimal traffic assignment, toll, and incentive scheme are under development for future work.

The Berkeley Integrated System for TRansportation Optimization (BISTRO)In chapter 4, I present a methodological framework for regional-scale multi-objective optimization of transportation systems using agent-based simulation (ABS), called the Berkeley Integrated System for TRansportation Optimization (BISTRO). BISTRO is an open source transportation planning decision support system that enables the simultaneous optimization of multiple transportation system interventions, including adjustment to public transit operations and vehicle fleet mixes as well as pricing strategies, using state-of-the-art activity- based travel models and ABS to explore the individual- and system-level impacts of packages of TDM strategies. In contrast to the state of the practice in transportation planning of analyzing a select few strategies across a discrete set of scenarios, the simulation-based multi- objective optimization approach that I explore in this dissertation using BISTRO leverages the machine learning principles of data exploration and exploitation to explore the interactions of policy design parameters and their outcomes to a much greater extent, thereby producing richer insights about the optimal trade-offs between the various competing objectives of policy design and implementation. BISTRO was developed in partnership with researchers at the Institute of Transportation Studies at UC Berkeley and Lawrence Berkeley National Laboratory as well as a project team at Uber, Inc. I led the development of the objective function design for BISTRO by translating municipal and regional transportation goals into quantitative optimization objectives with consideration for social, environmental, and eco- nomic sustainability implications and testing the efficacy with which these key performance indicators (KPIs) functioned when applied in simulation-based multi-objective optimization. In addition, I led the design of the output analysis and visualization platform for BISTRO.

My individual contributions are presented in the Scoring function design, Inputs, and Output analysis and visualization sections of chapter 4 in addition to the corresponding sections of Appendix A. Chapter 4 also presents the results of a pilot study of BISTRO that was conducted as a machine learning competition hosted within Uber Technologies, Inc. The key lessons learned from the design and execution of this pilot study demonstrated BISTRO’s utility as a human-in-the-loop cyber-physical system: one that uses scenario-based optimization algorithms as a feedback mechanism to assist urban planners with iteratively refining objective function and constraint specification on multi-faceted intervention strategies (e.g., pricing, public transit scheduling, vehicle fleet mix). The remainder of my dissertation research focuses on the application of BISTRO for the design of congestion pricing strategies with mode-specific incentives.

Congestion Pricing Optimization Case Study I: Sioux FauxThe fifth chapter of this dissertation presents a case study of the optimization of congestion pricing policy design using BISTRO and the Sioux Faux benchmark model presented in Chapter 4. The study exemplifies how the granularity offered by activity-based travel models can be leveraged to enhance the interpretability of multi-objective transportation policy optimization through rich analyses of the effects of policy design on both individual- and system-level outcomes. The location and size of a circular charging zone as well as two different pricing schemes (a cordon fee and a cordoned mileage fee) were encoded as inputs to an ABS with an activity based travel model of 15,000 travelers. Through an analysis of the effects of various weighting schemes across KPIs representing congestion- (i.e., total VMT, average vehicle hours of delay (VD) per passenger trip, and total GHG emissions), social- (i.e., average generalized travel cost burden of work and secondary trips), and revenue-based (i.e., total toll revenue) objectives, I developed a method for interpretation of the inherent trade-offs in transportation policy optimization and demonstrate the importance of cultivating transparency in policy decision support systems that use black-box optimization in order to produce explainable, defensible policy strategies.

I find that cordoned mileage fees Pareto-dominate cordon tolls, meaning they produce greater improvements across all objectives studied. I estimated an empirical 3-dimensional Laffer curve for each pricing strategy, representing the concave relationship between the main components of the pricing scheme design - the average toll paid per driving/TNC trip and the coverage of the cordon (in % of trips affected) - with the toll revenue and the driving/TNC mode share. I analyzed the relationships between these design parameters and the other KPIs, demonstrating the utility of simulation-based multi-objective optimization for human- in-the-loop transportation policy design. I find that the prioritization of congestion reduction poses a challenge for congestion pricing optimization in that it may result in unnecessarily large mode shifts away from driving. This significantly worsens the cost burden of travel (i.e., the total generalized cost of travel - including both travel cost and VoT-weighted travel time - as a portion of individual income) by disproportionately increasing the travel times of lower income individuals in a manner that may arguably be incommensurate with the benefits of congestion mitigation. Finally, I explore the role of weighting schemes in shifting the priority of an optimal congestion pricing scheme to social equity while maintaining congestion mitigation and toll revenue, the latter of which may be used to further improve the transportation system.

Congestion Pricing Optimization Case Study II: The San Francisco Bay AreaIn chapter 6, I build upon the Sioux Faux case study by designing a Pareto-based optimization of a cordoned mileage fee and incentive scheme for the City and County of San Francisco using BISTRO. I developed and calibrated an activity-based model of 50,000 commuters in the San Francisco Bay Area for the case study, including both ride alone and pooled on-demand rides in 7-seater vehicles and a multinomial logit mode choice model with coefficients estimated from a general population stated preference survey of San Francisco Bay Area residents that I conducted in 2018. I applied the Covariance Matrix Adaptation Evolution Strategy (CMA-ES) algorithm to optimize the hypervolume of the Pareto frontier of various transportation system objectives, including the minimization of the total VMT, average VD per passenger trip, total GHG and PM2.5 emissions, average cost burden, and net public revenue resulting from the congestion pricing and incentive scheme. Rather than produce a single ’optimal’ policy design, the approach taken in this study produces rich insights regarding the correlations between policy design parameters and the corresponding outcomes and produces a set of policy design options representing the optimal trade-offs between policy objectives.

I find that the inclusion of monetary incentives for pooled TNCs and public transit in a congestion pricing scheme for San Francisco improves all KPIs under consideration except for the net public revenue produced by the scheme, which may still be improved with respect to a baseline with neither tolls nor incentives. Incentives serve to amplify the operational and environmental benefits of congestion pricing by offering an additional positive financial incentive for the use of pooled modes of transportation (in addition to the negative incentive posed by the congestion charge), while reducing the average cost burden of travel across the region. This is particularly important in the San Francisco Bay Area where a majority of commuters traveling in the City of San Francisco either live or work in one of the other eight counties of the region where there is comparatively poor access to alternatives to driving. Incentives aid in compensating the most cost-sensitive drivers for the considerable travel time or cost increases they experience when shifting to public transit or pooled on-demand rides, respectively. This study demonstrates the value of Pareto-based optimization using ABS for developing dynamic TDM strategies capable of adapting to ever-changing travel behavior and policy priorities. I conclude this chapter with a discussion of actionable insights regarding the potential of congestion pricing and incentive schemes to achieve a variety of policy objectives, particularly in the presence of high-capacity pooling services such as microtransit.

Summary of Key Findings and Research DirectionsThis dissertation develops theoretical, methodological, and empirical contributions to the field of transportation engineering, specifically for the design and optimization of sustainable and equitable congestion pricing strategies. The multi-objective simulation-based optimization framework I have developed and applied using BISTRO lays the foundation for the continued development of algorithmic transportation policy decision support tools.

The use of activity-based travel models and ABS enables the examination of the network effects of pricing strategies, which are particularly relevant when considering 1) the regional impacts of pricing on mode shifts, since behavioral responses in one sub-region can influence those of another through the secondary effects of the pricing strategy on travel times and the level of service of alternative modes (e.g., reliability of public transit or on-demand ride services), 2) on-demand vehicle fleet operations, which are designed to optimize the revenue generated from the strategic positioning and dispatching of vehicles across a market, and 3) the economies of scale of pooling, which generate a virtuous cycle of higher pooled ride match rates, better pooled ride service reliability and affordability, and reduced costs with respect to greater density of pooled rides in any given area and time period. This dissertation presents the first multi-objective congestion pricing and incentive scheme optimization studies using ABS with an activity-based travel model that includes numerous travel options to access/egress public transit and, in the second case study, includes pooled on-demand ride services with survey-based estimates of the sensitivity of demand for pooling incorporated into the behavioral dynamics of the model.

This research has produced applicable insights for both the immediate- and long-term. While congestion pricing policies are currently under development in cities across the globe, there is also a broad trend toward ’smart city’ technologies that has been underway for over a decade, with increasing integration of sensors and connected control mechanisms to automate the operation of certain civil systems such as water, energy, and traffic management. In addition to updating traffic control systems (i.e., automated ramp metering, traffic lights, etc.), transportation agencies are integrating fare payment across transportation services and facilities and considering mobility as a service (MaaS) models that bundle transportation services and enable the seamless allocation of credits and incentives to nudge travel behavior.

The results of the two congestion pricing optimization case studies (chapters 5 and 6) con- tribute valuable insights for the ongoing development of transportation pricing strategies such as congestion pricing and MaaS. The San Francisco Bay Area case study in chapter 6 suggests that low-income incentive schemes can drastically improve the efficacy of congestion pricing to achieve congestion mitigation and environmental emissions reductions objectives while also reducing the average cost burden of travel. These benefits come at the cost of reduced public revenues that could otherwise be used to expand and improve public transit access. Targeted public transit improvements may also serve to reduce the cost burden of congestion pricing schemes by reducing the travel time for individuals who shift away from driving due to the congestion charges. However, in a region as large and varied as the San Francisco Bay Area, strategizing which investments to make and where to make them is a challenging task. The optimization of targeted reinvestment of congestion charging revenue into public transit is an area ripe for future work using BISTRO. As was demonstrated in the pilot BISTRO study (chapter 4), BISTRO can be used to optimize the public transit vehicle fleet mix, route schedules and headways, and fares in addition to the pricing parameters optimized in the congestion pricing case studies.

In addition to generating practical insights for the design of congestion pricing schemes, this dissertation demonstrates the challenges and opportunities for algorithmic policy design. Building upon the lessons learned from the BISTRO pilot study presented in chapter 4, the methodological approach developed in chapters 5 and 6 focuses on improving the explainability and interpretability of simulation-based multi-objective optimization by using BISTRO to explore the various aspects of the transportation system optimization problem, including the:• design of policy input parameters and an appropriate search space that is both large enough to enable exploration of potentially unexpected solutions and constrained enough to avoid the unnecessary use of resources exploring extreme solutions with undesirable outcomes, • specification of KPIs that are both mathematically sound and representative of realistic policy objectives (spanning congestion mitigation, equity, and financial feasibility) with which to guide the optimization algorithm to find insightful solutions, • development of a realistic and meaningful travel model that includes key behavioral, physical, and operational dynamics relevant to the optimization problem at hand (e.g., configuring the mix of fuel consumption and seating capacity in the public transit and on-demand vehicle fleets), • implications of scalarization schemes and the impacts of weighting conflicting policy objectives on the results of optimization, and • application of nuanced analysis to interpret and communicate the implications of the optimization results at both the individual- and system-levels.

The research presented in this dissertation offers a launching point for future work in the field of transportation policy design and optimization. BISTRO offers the capability to expand the scope of the congestion pricing optimization problem I have studied thus far to include aspects such as induced demand, vehicle-based charging, dynamic pricing structures, parking pricing, and public transit investment. In addition, BISTRO may be used as a tool to develop methodological approaches for robust transportation system optimization that produces strategies that perform optimally across a variety of scenarios. Such scenario-based optimization may be conducted across multiple baselines (as opposed to a single BAU), each representing a potential variation in the distribution of land use, mobility preferences, or other factors that may affect the distribution of transportation supply or demand. BISTRO can also be used to develop algorithms for transfer learning of policies across varying scenarios and geographies. Such methodologies would characterize hyper-parameters representing factors in the transportation network, services, and demand profile that are significant across scenarios or geographies, thereby increasing the efficiency and robustness of the optimization.

Lastly, a policy optimization framework relies heavily on the design of the objective function. The case studies in this dissertation applied just one of the three alternative social choice functions presented in chapter 3. More work is needed to continue developing an equitable approach for defining and operationalizing the notion of social optimality with respect to both equity and sustainability. The level of service KPIs implemented in BISTRO, including measures of public transit vehicle crowding, transportation cost burden, and accessibility are a first step in this regard, and the research I have conducted thus far on the impacts of weighted and Pareto-based optimization have illuminated the challenges in automating public policy decision-making.