Updating Regional Transportation Planning and Modeling Tools to Address Impacts of Connected and Automated Vehicles, Volume 2: Guidance (2018)

Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

30 Forward-looking planning activities typically seek to evaluate and choose from among a set of candidate decision options. For transpor- tation planners, this can include choices about expanding highway capacity, extending or enhancing transit service, creating bike lanes, and prioritizing system upgrades and maintenance. Planners often use models, along with expert judgment and other techniques, to assist in the evaluation and selection of decision options. The task is diffi- cult because there may be many and sometimes competing criteria by which the options must be evaluated, such as the safety, transportation demand, service quality, equity, greenhouse gas emissions, local air pol- lution, and cost. Moreover, how each option performs for those criteria depends on very hard-to-predict and often-disputed future conditions. These can include long-term demographic and land use changes, eco- nomic growth or decline, energy prices, consumer behavior, and new transportation technologies. Dewar and Wachs (2008) noted that transportation planning does not manage uncertainty particularly well: Travel demand forecasting as widely practiced today deals inadequately with uncertainty. . . . The current transportation modeling process is demanding in the sense that it employs a great deal of data to a large number of interconnected models having many parameters. The com- plexity of the modeling process, however, does not extend to the accurate representation of complex economic and social phenomena, and point estimates of many quantities are used that make it difficult to analyze or even to represent the uncertainty that characterizes trans- portation systems and traveler decision making. Uncertainty in Transportation Systems The extent to which uncertainty exists in the performance of transportation systems and in systematic responses to uncertainty in those systems through the planning, decision making, operations, and management processes is an important and complex problem. Urban trans- portation systems include extensive networks of massive, immovable, and long-lived physical facilities. The extent, location, and physical condition of the current system in any geographic location are, in the short run, among the least uncertain of all the elements of the physical envi- ronment. Bridges, tunnels, highways, and rail lines are unchanging for decades or centuries, are dominating features of the landscape, and are difficult to alter physically. What is important to structure the discussion here is not the stability of the physical trans- portation network but rather the variability of travel on that network and, consequently, the C H A P T E R 5 Planning in the Context of Uncertainty Chapter Highlights • Reviews how uncertainty is managed in current transportation planning. • Describes the unique challenges in managing uncertainty posed by CAVs. • Identifies methods suited to man- aging decision making under deep uncertainty: – Scenario planning, – Assumption-based planning, – Robust decision making, – Info-gap, and – Dynamic adaptive pathways planning.

Planning in the Context of Uncertainty 31 variability and uncertainty of the network’s performance under differ- ent circumstances. Travel that takes place on transportation facilities is highly variable, flexible, and malleable. People and goods use the transportation system rationally, but they employ many and highly individual criteria when deciding how to fulfill varying needs. The com- plexity of travel decision making by people and firms is fundamentally reflective of social, economic, and cultural patterns that are themselves quite complex, and these are compounded by the complexity of physi- cal flows in transportation networks. Yet, when society taken together makes all of its travel decisions by using many different rational choice processes, the outcome is clear patterns that seem regular and repetitive, and this in turn leads to the notion that uncertainty is less important to planning than it actually is. Traffic peaks almost every day at the same times and places; roughly the same number of people use public transit versus highways between certain origins and destinations at a certain hour of the day. When looked at by an engineer, traffic on a facility has certain predictable characteristics like volumes, densities, starting times, and concentrations at certain origins and destinations that recur on a predictable, daily basis. However, the engineer looks at the performance of the system and not of the thousands of people who are using it. When looked at as a social phenomenon rather than as traffic flows, trips can be made by different modes, at different times, at different vehicle occupancy rates, for different purposes, from different origins to different destinations, and, in at least some cases, can be postponed or cancelled. Transportation modelers are no strangers to uncertainty, although it is typically left unad- dressed. CAVs exacerbate many of the existing uncertainties, such as the cost of driving, the elas- ticity of travel demand to this cost, and the greenhouse gas emissions per mile traveled. If these aspects were previously treated as uncertain, the range of potential values for these parameters might need to be widened, and if they were previously not treated as uncertain, they certainly must be now. There is also a host of other ways that CAVs create new challenges for managing uncer- tainty. AVs in particular can be thought of as a new transportation mode and may funda- mentally change the future of mobility and its associated effects (DOE 2016). As just one example, AVs are expected to fundamentally change land use in the decades to come, not only because people’s preferences for urban versus suburban living may change, but because significant changes in urban environments may occur. Urban dwellers may give up their per- sonally owned vehicles in favor of SAV services, which means the large fraction of urban space devoted to parking may be converted to other uses (Zmud et al. 2016). Housing shortages in many cities could be alleviated if garage space previously devoted to vehicles could be repur- posed for housing. If AVs prove to be extremely safe and efficient, redesigned lanes may free up space for other modes. These types of changes are enormously difficult to anticipate. Just as it was not possible to predict in 1990 how the Internet would change communication methods and frequency 20 years later, so it is not possible today to confidently predict how AVs will change travel modes and frequency 20 years from now. Yet, the uncertain future will shape the answers to questions that transportation planners are trying to address now. Should a particular light rail line be expanded, or will that expansion become obsolete in a future with AVs? Should urban planners consider purchasing satellite parking for AV fleets? Should an AV lane be included in future highway capacity expansions? These questions are difficult to answer with traditional methods of planning. The next section examines how these questions can be addressed with a new class of methods for managing deep uncertainty. Because of the complexity in the modeling process, point estimates are used, which makes it difficult to repre- sent the uncertainty that characterizes transportation systems and traveler decisions.

32 Updating Regional Transportation Planning and Modeling Tools to Address Impacts of Connected and Automated Vehicles Overview of Planning Processes Predictive Analysis, or Agreeing on Assumptions Most traditional planning methods seek to (a) reduce uncertainty by requiring agreement on assumptions about the current and future conditions under which a plan must perform, and (b) analyze the decision options. Transportation planners would first characterize the future urban form, economic growth, and other factors that affect travel demand. These characteriza- tions are often, but need not be, single numerical values; they could also be distributions around future trends. For example, instead of predicting an increase in VMT of 10%, a planner could predict a normally distributed increase in VMT with a mean of 10% and a standard deviation of 2%. In this case, a Monte Carlo simulation would be used to estimate the most likely future given the assumptions that were adopted, and to identify the near-term policy actions that would maximize the likelihood of the desired outcomes (Adler et al. 2014). Transportation planners would then evaluate the merits of various plans or investment choices (e.g., highway capacity expansion) under these assumptions. A sensitivity analysis could help assess how much influence each assumption has on the outcome. Such approaches have been termed “agree-on-assumptions” (Kalra et al. 2014), “predict-then-act” (Lempert et al. 2013), or “science first” (Dessai and Hulme 2007). When faced with disagreement and deep uncertainty (e.g., about the deployment and impact of AV technology), these traditional processes are vulnerable to bias and gridlock. First, many important assumptions are buried in models rather than in front of decision makers. This makes it difficult for decision makers to understand and assess potentially critical assumptions on which their decisions hinge. Second, many factors are difficult, if not impossible, to predict. Stakeholders also know that the choice of assumptions drives the choice of investment option. They may press for assumptions that will lead to the options they already favor and thereby make consensus difficult (Lempert et al. 2003). Decision makers risk losing stakeholders’ buy-in early if the foundations of the decision process lack transparency, appear arbitrary, or do not include their beliefs. Agree-on-assumption approaches are also vulnerable to reaching brittle decisions—ones that are optimal for a particular set of assumptions but that perform poorly or even disastrously under other assumptions. Sensitivity analyses are often not sufficient for exploring the full range of plausible assumptions and future conditions (Bonzanigo and Kalra 2014), and agree-on- assumptions create little opportunity for exploring the performance of decision options under unexpected conditions. They yield no information about how an optimal solution performs if the future is surprising, and they do not guide decision makers to solutions that might work well if the predicted future does not happen. Yet, a significant need for understanding the effect of surprises and unexpected conditions exists; repeated studies have shown that human beings have a widespread tendency toward overconfidence, with strong belief in our ability to predict the future when we cannot (Kahneman 2011). Exploratory Analysis or Agreeing on Decisions It is possible to manage deep uncertainty by seeking a robust decision—one that performs well across a wide range of futures, preferences, and worldviews, though it may not be optimal in any particular one. Robust decisions are often flexible—designed to be modified over time as new information becomes available. It is possible to identify robust strategies by inverting the traditional steps (i.e., by using agree-on-assumption processes). These strategies, also sometimes called “context-first” methods (Ranger et al. 2010), begin with laying out the decision options (as opposed to first laying out predictions of the future)

Planning in the Context of Uncertainty 33 and then stress-test the options under a wide range of plausible conditions, without requir- ing a decision or agreement upon which conditions are more or less likely. They evaluate the decision options repeatedly, under many different sets of assumptions. Planners can evalu- ate options under low-likelihood but high-consequence events, can treat as uncertain the assumptions buried in models, and can use every stakeholder’s beliefs about the future; agree- ment on assumptions is not required. This process reveals which of the options are robust, meeting needs under a wide range of conditions rather than performing well in only a few. Analytical tools can then help identify the specific conditions in which each option no longer meets its goals. Analyses performed in this way do not make the decisions for decision makers. Instead, they help decision makers debate important questions: • Are the conditions under which our option performs poorly sufficiently likely that we should choose a different option? • What trade-offs do we wish to make between robustness and, for example, cost? • Which options leave us with the most flexibility to respond to changes in the future? This inverted process promotes consensus around decisions and can help manage deep uncer- tainty around transportation technology, climate change, and a host of other factors. Qualitative Methods for Managing Deep Uncertainty The most common qualitative method for addressing uncertainty is scenario planning. Scenario planning is a response to the limits of pre- dictive agree-on-assumptions analysis. In scenario planning, analysts develop diverse and often divergent views about the long-term future. Rather than tell a single story, the planners craft a suite of different tales. Classical scenario planning involves a small set of handcrafted scenarios aimed at facilitating broader thinking about potential future outcomes. Yet, it has limitations in ensuring that a wide enough range of futures is considered and in linking scenarios to near-term policy choices. A newer group of methods has evolved in response. These methods typically use vast numbers of computer-generated scenarios to identify actions that are robust in performing well across a wide range of potential futures. Classical Scenario Planning Scenario planning constructs diverse and often divergent narratives about the long-term future. A family of scenarios, often three or four, aims to span the range of plausible futures relevant to the decision at hand. The aim is for planners to use those scenarios to consider how near-term policies might shape and be shaped by those futures. Nearly all the work described earlier uses scenario planning to assess CAV impacts. In addition, FHWA is pursuing a study to develop scenarios for CAVs. The Task Order Proposal Request describes the study (FHWA 2016) as follows: The purpose of this study will use the transportation scenario planning process to develop approximately three to five descriptive futures (scenarios) of the deployment, market uptake, use, and impacts of CV and AV technologies. The U.S. Department of Transportation (DOT) is planning to provide the outcome of this study to State, regional and local transporta- tion agencies. The deliverables of this study shall include the future scenario outcomes, a high level assessment of these futures, and an illustration of how agencies can use scenario Qualitative methods include scenario planning and assumption-based plan- ning. The former has limitations in the range of futures considered and in linking to near-term policy choices. Thus, the latter has evolved in response.

34 Updating Regional Transportation Planning and Modeling Tools to Address Impacts of Connected and Automated Vehicles planning to develop their own, more localized future CAV scenarios. State and regional agencies may use this illustrative scenario planning process to anticipate likely issues and challenges they will face due to CAV adoption, and therefore to help visualize and under- stand their planning options, including developing or changing institutional and operational responses and policies. The Department of Energy also recently engaged in a scenario planning exercise to identify trends that would lead to different greenhouse gas emission scenarios in transportation. Mem- bers of this team were part of workshops that resulted in the paper The Transforming Mobility Ecosystem: Enabling an Energy Efficient Future, which has several goals (DOE 2016): The intent of this paper is to provide the Energy Department’s forethought, along with public and private stakeholder input, on the future of mobility and subsequent impacts on energy. It introduces four possible mobility futures, or narratives, defined by two factors chosen for their transformative potential and pertinence to the discussion: vehicle control (from driver-only, to self-driving, and fully automated defined as Level 4 or 5 functionality by SAE International), and vehicle owner ship (from personal ownership to fully shared vehicles). This paper, however, does not offer policy recommendations, or provide strategies that would enable one future narrative over the other. Additionally, it does not set target expectations for low-carbon technologies (e.g., battery costs), but observes a range of factors that, ultimately, could shape each future narrative and determine the impacts on energy and GHG emissions. Other factors may emerge in the future that could have a significant energy impact. Scenario planning does have shortcomings. First, the choice of any small number of scenarios to span a highly complex future is ultimately arbitrary. A scenario exercise will inevitably miss many important futures that do not make the cut into the top few. Second, scenario-based planning provides no systematic means to compare alternative policy choices. With a small set of divergent scenarios, it can be unclear which if any scenario to use for plan- ning purposes, and how to choose from among competing policy options that make sense in some but not all the scenarios. Thus, scenario planning is a powerful tool for imagining the uncertain future and constructing collective ideas of desirable and undesirable futures, but these ideas may be difficult to use in decision making to shape those futures (Lempert et al. 2003). Assumption-Based Planning Assumption-based planning is another qualitative approach to managing uncertainty. All plans must make assumptions about the future because of the presence of uncertainty. Many assumptions are explicitly identified and planned for, but most plans also contain implicit or hidden assumptions. These implicit assumptions can cause significant problems and weaken robustness of plans when not made part of the planning process. In the presence of uncertainty, assumption-based planning (Dewar 2002) can help organizations systematically identify explicit and implicit assumptions and ensure they are part of the process of planning actions to achieve defined strategic goals. Assumption-based planning hinges on identifying load-bearing assumptions (assump- tions that, if broken, would require major revision of the course of action) and, subsequently, the vulnerability of those load-bearing assumptions. For example, a plan to expand light rail to a planned development area hinges on the assumption that the proposed development will succeed. This assumption might be vulnerable if the development hinges on an optimistic level of economic growth. Assumption-based planning guides organizations in determining a course of action to deal with the vulnerability of load-bearing assumptions once they are identified. An MPO in a high- growth area with sig- nificant congestion and investment needs may implement scenario planning and include CAV and other techno- logical impacts in the long-range plan.

Planning in the Context of Uncertainty 35 Quantitative Methods for Managing Deep Uncertainty In response to the difficulty of linking scenarios to policy choices, many have turned to alter- native methods for decision making under deep uncertainty. As described earlier, deep uncer- tainty exists when decision makers do not know or do not agree on the models that describe relationships between key drivers and outcomes, the probabilities of key variables, or how to value the desirability of different outcomes. These characteristics hold true of CAV technologies, for which it is not clear how the different technological, social, economic, and other trends will interact to shape CAV adoption or their outcomes. A new class of methods—which includes but is not limited to RDM (Lempert et al. 2003), dynamic adaptive policy pathways (Haasnoot et al. 2013), and info-gap (Ben-Haim 2006)—all seek to use multiple (often hundreds or thousands) of computer-generated future scenarios to identify decisions that are robust—that is, decisions that work well in many future scenarios even if they are not optimal in any one future. Such decisions are often no regret (i.e., they make sense no matter what the future brings) and adaptive or flexible (i.e., they enable changes as new information becomes available). These methods and bodies of work have been brought together by a recently formed professional organization called the Society for Decision Making Under Deep Uncertainty (www.deepuncertainty.org), which updates its website with publications and reports describing these methods. As the website shows, these approaches have been widely applied to water resource planning, defense planning, energy investments, health, and a variety of other fields but have not been extensively applied to transportation planning. There is clear value in doing so, and it is to be hoped that the deep uncertainties that CAVs present will encourage such work. Robust Decision Making RDM rests on a simple concept (Lempert et al. 2003). Rather than use models and data to assess decision options under a single set of assumptions, RDM runs models over hundreds to thousands of different sets of assumptions to describe how plans perform in many plausible conditions. Unlike Monte Carlo analysis, which attaches probabilities to those assumptions to estimate expected outcomes, RDM uses simulations to stress-test strategies and helps decision makers identify robust strategies—those that perform well regardless of the assumptions or future conditions—and identify the key trade-offs between potential robust strategies. Often, the robust strategies identified by RDM are adaptive, that is, designed to evolve over time in response to new information. Info-Gap Theory Info-gap theory is another approach that helps decision makers identify robust options, but it takes a somewhat different tack. RDM uses models to assess the performance of options in a wide range of potential future conditions and then identify conditions that result in poor per- formance (i.e., conditions to which the system is vulnerable). In contrast, info-gap uses models to compute how options perform as a function of uncertainty. An info-gap analysis defines robustness as “the maximum uncertainty in our estimates that can be tolerated while still guar- anteeing a particular desired result” (Irias and Cicala 2013). An info-gap analysis produces a graph showing the performance that planners can robustly achieve on one axis as a function of uncertainty on the other axis. Like RDM, info-gap does not provide decision makers with the solution; rather, it seeks to inform decision makers on trade-offs, risks, and vulnerabilities. Quantitative methods include robust decision making, info-gap, and dynamic adaptive pathways. Rather than ask, “What will happen?” these methods ask, “What should we do today to most effectively manage the range of events that might happen?”

36 Updating Regional Transportation Planning and Modeling Tools to Address Impacts of Connected and Automated Vehicles Dynamic Adaptive Pathways Planning Another approach is dynamic adaptive pathways planning (DAPP; Haasnoot et al. 2013). With the DAPP approach, a plan is conceptualized as a series of actions over time (pathways). The essence is proactive planning for flexible adaptation over time, in response to how the future actually unfolds. The DAPP approach starts from the premise that policies and decisions have a design life and might fail as the operating conditions change (Kwadijk et al. 2010). Once actions fail, additional or other actions are needed to achieve objectives, and a series of path- ways emerges; at predetermined trigger points, the course can change while the objectives are still achieved. By exploring different pathways and considering the path-dependency of actions, planners can design an adaptive plan that includes short-term actions and long-term options. The plan is monitored for signals that indicate when the next step of a pathway should be imple- mented or whether reassessment of the plan is needed. Institutional, Resource, and Other Considerations for Managing Deep Uncertainty While scenario planning and other qualitative methods for managing uncertainty have been around for many decades, quantitative methods may be particularly unfamiliar, since they have been enabled only recently by the growing availability of computational resources. Thus, insti- tutional challenges to mainstreaming new approaches exist, even if they may be better suited to a particular analytical problem. This section presents some of those considerations. These observations are adapted from Lempert et al.’s (2003) comments on the application of RDM in developing country contexts. Quantitative methods are generally designed to employ existing models and data. Thus, in cases where decision makers are already using quantitative analysis to inform their choices, these methods can augment such activities to provide a richer understanding of uncertainty and the best ways to respond to it. The models used in these analyses can be simple or com- plex. For instance, an analyst using a simple spreadsheet model to compare the cost–benefit ratios of alternative investments could use these methods to run the spreadsheet over many thousands of combinations of assumptions and to identify those futures where one invest- ment is consistently more cost effective than another. Analysts with a large, complex model could similarly use these quantitative methods to stress-test the strategies that emerge from their analysis. As one potential implementation barrier, compared with a traditional approach, quantitative methods in particular require more computer processor time to conduct hundreds to thousands of runs and more computer storage to save the results. In practice, these are not significant constraints. Analysts with spreadsheet models will generally have more than sufficient storage and processing power on a laptop to run the spreadsheet thousands of times. Analysts running a complicated model may require hundreds or thousands of processors to run their models over numerous cases. These are increasingly available (for instance, Amazon now rents time on its huge stock of multiprocessors), and those with the skills to build complicated models can also access such multiprocessor systems. Configuring a model to run hundreds to thousands of cases often represents the greater chal- lenge. For instance, staff skilled at developing cost–benefit spreadsheets may not know how to run the spreadsheet automatically for thousands of cases. Complex models may have an input file structure that makes it difficult to run thousands of cases efficiently. Both situations may require training and some reworking of computer code to enable analysts to generate and batch runs. Fortunately, this software, along with related training, proves to be a sound investment because it is generally useful for a wide range of analyses.

Planning in the Context of Uncertainty 37 Perhaps the most significant challenges to implementing quantitative methods of managing uncertainty arise because these methods represent a new way of thinking about how near-term actions can best manage future risks. Analysts are generally trained in predictive thinking, and the decision makers they inform often expect predictive quantitative information. Methods of managing uncertainty answer a fundamentally different question. Rather than ask, “What will happen?” they allow analysts and decision makers to ask, “What should we do today to most effectively manage the full range of events that might happen?” Using these methods requires training for analysts and a path by which organizations become comfortable using new and more effective types of quantitative information. One successful path involves conducting a demon- stration project parallel to an organization’s regular planning activities. Once the demonstration is complete, the organization can use this experience to begin to fold the new methods into its planning.