Dynamic Fleet Management for Autonomous Vehicles: Learning- and optimization-based strategies

Autonomous vehicles (AVs) have been heralded as the key to unlock a shared mobility future where transportation is more efficient, convenient, and cheaper. However, the AV utopia can only come to fruition if the majority of users trust that autonomous mobility-on-demand (AMoD) systems are on a par with owning a vehicle in terms of service quality. Once the perception of quality is highly subjective, we propose a more personalized approach to on-demand mobility, in which users are segmented into service quality classes. These classes comprise minimum requirements regarding responsiveness and privacy, allowing us to model a series of user profiles formalized using strict service quality contracts. By honoring these contracts, providers can build users' trust and gain their loyalty, which on a grander scheme can contribute to a faster transition to a shared mobility future.

This thesis presents a series of strategies to guaranteeing service quality throughout operational scenarios arising in the timeline of AV technology deployment. First, a precondition to providing service quality in autonomous transportation is safety. During a transition phase to full automation, AV operation will likely be restricted to areas where safe operations are guaranteed, leading to the formation of hybrid street networks comprised of autonomous and non-autonomous vehicle zones. In this setting, meeting user service quality expectations is primarily a matter of coverage, once mobility services will have to access both AV-ready and not AV-ready areas. Accordingly, this thesis proposes solutions to overcome the challenges entailed by such a transition scenario, where infrastructures, regulatory measures, and AV technology are gradually evolving.

Then, assuming that widespread automated driving is the new status quo, we set out to model rich autonomous transportation scenarios comprised of heterogeneous users and vehicles. Central to our analysis is finding an adequate tradeoff between fleet size and service quality. In traditional AMoD systems, providers can do only so much to prevent user dissatisfaction since, to some extent, this is a matter of having enough vehicles. When the demand outstrips the supply, users inevitably experience longer delays or even rejections, ultimately undermining trust in the service. However, these shortcomings may plague future transportation systems only if setting the fleet size and mix remains a strategic decision. In contrast to most related literature, this thesis investigates a disseminated AV ownership scenario, where ridesharing platforms can occasionally hire available privately-owned AVs on-demand. In this scenario, customers can simultaneously own and share AVs, a setup that better resembles the operation of today's transportation network companies (TNCs), which rely entirely on micro-operators. As a result, AMoD systems can increase and decrease vehicle supply in the short term, thus shifting fleet sizing to the operational planning level.

Moreover, analogously to other transportation modes, we consider that the system must deal with a diversified user base with different service quality expectations. This setup allows providers greater leeway to explore requests' delay tolerances to design efficient routes. To balance user expectations and avoid an oversupply of vehicles, we propose a multi-objective matheuristic that dynamically hires third-party AVs to meet the demand. Our approach adds to recent literature by allowing providers to prioritize different customer segments, besides choosing the exact tradeoff between meeting each segment's needs and hiring extra vehicles. This way, when vehicles are lacking, the optimization process can steer the ride-matching solution towards addressing user requests in order of importance (e.g., most lucrative first). To make the most of currently working vehicles, we also design a repositioning algorithm that fixes supply and demand imbalances using users' service level violations as stimuli.

Further, to enable anticipatory decision making, this thesis incorporates the stochastic information surrounding both privately-owned AV supply and heterogeneous passenger demand in the fleet management process. We propose a learning-based optimization approach that uses the underlying assignment problem's dual variables to iteratively approximate the marginal value of vehicles at each time and location under different availability settings. In turn, these approximations are used in the optimization problem's objective function to weigh the downstream impact of dispatching, rebalancing, and occasionally hiring vehicles. By harnessing the historical knowledge regarding both demand and supply patterns, we show that AMoD providers are substantially better equipped to meet user needs without necessarily having to own large AV fleets.

Typically, learning-based fleet management strategies end up reinforcing biases present in the demand data, therefore frequently moving towards cities' most affluent and densely populated areas, where alternative mobility choices already abound. Although lucrative for providers, this fleet management strategy runs counter to a broader city goal of equitably distributing accessibility across all regions and population demographics. To counterbalance the demand biases, we investigate the extent to which fare subsidization policies can drive the learning process towards sending vehicles to targeted regions where accessibility is lacking. Our results suggest that by using an adequate scheme of incentives, policymakers can orchestrate transportation providers to diminish the insidious effects of ``cream-skimming'' practices, thus using AVs in favor of mobility equity.

Lastly, once we have designed strategies that balance the goals of cities, independent owners, fleet owners, and users, we focus on a different approach to maximizing fleet productivity in urban environments. No matter how efficient a fleet optimization method can be, by limiting AVs to service a single commodity type (i.e., people), fleet utilization and consequently profits are bounded by passenger demand patterns. As autonomous technology evolves, however, new opportunities to improve asset utilization arise. We end this thesis with a model for a versatile transportation system where mixed-purpose compartmentalized AVs can address both passengers and goods simultaneously. With the growth of e-commerce and same-day deliveries, our approach provides a starting point to study more flexible short-haul integration systems to consolidate passenger and freight flows.