Oded Cats

Ride-hailing platforms that offer both private and pooled ride-hailing services must jointly determine how to allocate their available fleet between the two services and how to price each. These two decisions are inherently interdependent: rider choices between private and pooled ride-hailing depend jointly on the fares and expected service conditions of each option, while those conditions are in turn shaped by how available drivers are allocated across services. Notwithstanding, past studies, have largely optimized pricing and service-type assignment separately, considered only a single service type, or treated the vehicle service mix as fixed rather than as a decision that can be adjusted in real time as demand evolves. This study addresses this gap by proposing a cooperative multi-agent reinforcement learning framework that jointly and dynamically optimizes pricing and vehicle-level service-type assignment under endogenous, choice-based rider demand, modeled using a utility-based discrete-choice model. We formulate the problem as a multi-agent Markov Decision Process and train a pricing agent together with a service-type assignment agent using a cooperative PPO-based training algorithm within a purpose-built ride-hailing simulation environment. To scale to large, variable-size fleets without a corresponding increase in the number of model parameters, the service-type agent adopts a parameter-sharing architecture across vehicles. Experiments on the Delft and Amsterdam road networks, with fleets of up to 1,000 vehicles, show that the proposed framework learns stable, coordinated pricing and service-type assignment policies that achieve higher revenue and cumulative return than rule-based and single-component learning baselines, while remaining responsive to shifts in market scale and rider preferences. An ablation study further shows that parameter sharing reduces training time by roughly 22\% relative to a non-sharing variant.

Recent years have seen a push towards electrified public transport due to lower battery costs and regulatory requirements. However, full electrification remains constrained by battery capacities, weight constraints and capital costs of buses and charging infrastructure. This study considers Mobile Autonomous Charging Pods (MAPs) as a flexible and more cost-effective alternative for supporting high-frequency urban bus operations. It proposes an integrated simulation-based heuristic approach where a greedy heuristic optimizes bus and MAP battery capacities and the fleet size of MAPs, while a discrete-event simulation captures dynamic operations, including MAP dispatching, routing, energy constraints, and evaluates the feasibility of the proposed system. The approach is applied to real-world data from the Stockholm bus network. Our findings show that the MAP-based approach results in a reduction in total battery capacity and enables cost reductions of 47-52% when compared to depot-only solutions.

Urban flooding reduces transport infrastructure availability and weakens networks’ ability to serve travel demand. Prior studies examine individual modes or exposure-based disruption, with limited attention to whether affected demand can be served by alternative modes at acceptable service levels. This study develops a two-stage framework to assess unimodal robustness and cross-modal substitution in flood-disrupted multimodal networks. The first stage evaluates demand originally using car, bus, and metro that remains serviceable within its original mode under service and capacity constraints. The second stage reallocates flood-induced low-service unimodal demand across the multimodal network to quantify substitution ratios and modal contributions. Applied to Greater London, results show that low-service bus demand generates a large substituted volume, car demand is hardest to substitute, and metro demand has greater substitution potential. High-vulnerability areas show that flood vulnerability cannot be inferred from exposure or unimodal robustness alone; substitution depends on alternative modes’ coverage and residual capacity.

Travel time unreliability in public transport (PT) reduces user satisfaction, increases costs, and leads to revenue loss. Integrating pricing strategies with multi-modal development offers promising solutions to mitigate the adverse impacts of delays. This study extends the ”premium fare with delay insurance” concept, where passengers can choose to pay a surcharge that guarantees free access to alternative services if delays exceed a set threshold. We incorporate the demand change triggered by the introduction of the premium fare to an earlier version of the model. The model assumes that some users are captive to their current choice whereas some users of other travel modes may shift to the PT system in response to the availability of this new fare product. We hence utilize a dogit model formulation and apply it to the Washington DC metro with empirical delay and demand data. The model optimizes premium fare pricing to maximize operator profit while accounting for user behavior and compensation costs. The mean optimal premium fare level across all OD pairs is found to be around $4.35 compared to $2.52 for a standard ticket, with the delay qualifying threshold set at 5 minutes. The optimal fare levels are primarily driven by delay severity, distance and geographical characteristics. They vary across loyalty scenarios, with slightly lower mean optimal fares emerging when cross-mode demand response is considered. The results show that introducing premium tickets improves operator profitability and reduces passenger generalized travel costs, particularly by lowering travel time costs through induced mode shifts.

Railway operations are frequently affected by unplanned disturbances, causing timetable deviations and delay propagation. Real-time timetable rescheduling therefore requires rapid retiming and reordering decisions while satisfying interdependent operational constraints. This study proposes a Hierarchical Reinforcement Learning (HRL) framework that decomposes the rescheduling process into spatial--temporal subtasks. A high-level coordinator provides subgoals using an attention-based state representation, while a low-level executor makes retiming and reordering decisions based on detailed event information. The framework is integrated with an event-triggered mesoscopic simulation environment in which dynamic action masking excludes infeasible actions before execution, ensuring that the rescheduled timetable is conflict-free.The framework is evaluated on a busy railway corridor in the Netherlands. Numerical experiments assess the effects of action masking and hierarchical decision-making and compare the proposed method with flat reinforcement learning variants, First-Come-First-Served (FCFS), First-Schedule-First-Serve (FSFS), and a Mixed-Integer Linear Programming (MILP) formulation that provides optimal solutions for the considered instances. The ablation results show that hierarchical coordination yields more consistent performance than flat architectures, particularly under interacting disturbances. On the testing set, HRL reduces the average instance-wise optimality gap from 15\% for FCFS to 10\%, and the maximum gap from 58\% to 28\%. Its average computation time is 2~s per instance, compared with 26~s for MILP. These results demonstrate that the proposed framework achieves a favorable balance among solution quality, operational feasibility, and computational efficiency.