Sten Elling Tingstad Jacobsen, System- och reglerteknik

Electric Autonomous Mobility-on-Demand (E-AMoD) systems offer a path toward sustainable urban transportation through the coordinated operation of shared, zero-emission autonomous vehicles. Yet their deployment poses difficult operational challenges: fleet rebalancing, vehicle routing, and charging must be managed jointly, under uncertainty, and within tight computational budgets. A central argument of this thesis is that no single decision-making paradigm suffices. Optimization-based methods are well suited for strategic fleet control where uncertainty guarantees and feedback are essential, while learning-based methods become necessary at finer operational scales where real-time optimization is computationally prohibitive.
Three contributions are presented, each targeting a different operational level. The first introduces a chance-constrained model predictive control (MPC) framework for station-level fleet rebalancing, combining Gaussian Process Regression for probabilistic demand forecasting with a hierarchical architecture that separates strategic rebalancing from tactical matching. The second extends this framework to electric fleets operating under multiple interacting uncertainties, employing a tailored Nested Benders Decomposition to maintain metropolitan-scale tractability without sacrificing MPC's receding-horizon feedback. The third contribution shifts to node-level electric dial-a-ride routing, including pickup-delivery sequencing, time windows, and ride-time constraints, and proposes a deep reinforcement learning approach built on a Graph Edge Attention Network capable of handling hundreds of requests with second inference times. Taken together, the three contributions show that optimization and learning serve complementary roles in E-AMoD operations, with the appropriate paradigm determined by the granularity and real-time demands of the problem at hand.