From Forecasting to Scenario Planning: The Case of Autonomous Vehicles
MetadataShow full item record
CitationBauranov, Aleksandar. 2021. From Forecasting to Scenario Planning: The Case of Autonomous Vehicles. Doctoral dissertation, Harvard Graduate School of Design.
AbstractEfforts to forecast travel demand have led to the development of complex models which attempt to replicate human daily actions, choices, and movements. However, a growing body of literature suggests that the complexity of these models and their limited consideration of uncertainty have adversely affected their usefulness in the planning process. This dissertation argues that transportation planning should shift to methods that facilitate understanding and communication of uncertainty instead of relying on seemingly deterministic predictions of complex models. Two modeling paradigms – activity-based and scenario-based models are analyzed to show how they handle uncertainty in the case of assessing the travel impacts of autonomous vehicles.
Three metropolitan areas, Seattle, San Francisco, and the Detroit region, are used as case studies to estimate the impacts of autonomous vehicles on total travel and accessibility. The results of the activity-based modeling indicate that the effects of autonomous vehicles are different in different regions, primarily due to the differences in income, density, and access to public transit. While vehicle miles increase in all three regions, 17% in Seattle, 22% in the Bay Area, and 11% in Detroit, accessibility is highly dependent on the local context. The scenario-based model is not able to produce the results with this level of granularity. However, due to many unknowns associated with emerging technology, the scenario-based model proved to be better suited to incorporate various aspects of autonomous vehicles.
Beyond the estimates of travel impact, the results show that more informed planning can be achieved by moving away from deterministic forecasting and especially away from the urge to improve forecasting accuracy by building bigger models. Every piece of additional data and every additional parameter has an uncertainty cost that is compounded with the previous uncertainty costs. Instead, the modelers should aim to create more useful models by increasing the transparency of the modeling process and by reducing its complexity.
Citable link to this pagehttps://nrs.harvard.edu/URN-3:HUL.INSTREPOS:37370255