Urban waste collection in dense and spatially constrained city centers is shaped by more than the daily routing of vehicles. Collection feasibility and cost also depend on installed bin capacity, available vehicle types, route-duration limits, intermediate-facility unloading, and workload distribution across the fleet. Amsterdam city center motivates the problem setting as an illustrative example of this broader class of constrained urban logistics problems. The computational evidence in this thesis is therefore Amsterdam-inspired and benchmark-based, rather than calibrated to the actual Amsterdam collection system.
This thesis studies these interactions through an integrated optimization framework for municipal waste collection. The problem is formulated as a Waste Collection Multi-period Inventory Routing Problem with Intermediate Facilities and Bin Configuration decisions (WC-M-IRP-IF-BC). The model reverses the classical distribution-based interpretation of the Inventory Routing Problem by representing waste accumulation at customer nodes and preventing overflow beyond installed modular bin capacity. It jointly decides bin installation, visit timing, vehicle deployment, route sequencing, vehicle-load propagation, and intermediate-facility use over a finite planning horizon. The exact MILP uses a normalized scalar planning-cost objective that combines bin investment, vehicle deployment, routing distance, and, when activated, a distance-range workload-balance term.
The model is tested on modified Archetti benchmark instances adapted to the waste collection setting. The computational study evaluates a fixed source-capacity baseline, objective-cost ratio sensitivity, homogeneous and mixed fleet configurations under route-duration restrictions, workload-balanced solutions, and staged adaptation policies under demand growth. The fixed source-capacity baseline shows that integrated modular bin optimization reduces cost relative to an inherited storage layout in every comparable case with a usable solution, with improvements of 3.8\% to 18.9\%. Cost-sensitivity experiments further show that bin and deployment coefficients affect collection frequency, routing effort, and fleet use. Low bin costs encourage storage-intensive solutions, while high bin costs increase collection frequency and routing effort. Additional vehicles are most valuable when route-duration constraints are binding, and reduced-capacity mixed fleets can improve feasibility by splitting work across vehicles.
The workload-balance experiment substantially reduces distance-based workload gaps in several instances, but diagnostic comparisons show that distance balance does not necessarily imply compre hensive balance of driver burden. Under demand growth, reliability is best understood through adaptation pathways, where fixed plans may first absorb growth through routing effort, then require bin reallocation, bin expansion, fleet expansion, or combined expansion. Overall, the thesis shows that integrated strategic and operational planning provides a clearer basis for evaluating waste collection systems than routing cost alone. Within the controlled benchmark setting, the exact MILP formulation supports detailed structural analysis for small and medium instances, while larger networks, repeated visits to the same physical intermediate facility, stochastic demand, and richer workload measures motivate future scalable solution methods. ... Read more

Urban waterway networks offer a promising complement to land-based public transit, yet coordinating autonomous vessels with fixed bus timetables under dynamic passenger demand remains challenging. This thesis develops and evaluates a Multi-Agent Deep Reinforcement Learning (MADRL) framework for demand-responsive scheduling of autonomous surface vessels in an intermodal harbor setting. The thesis is positioned at the intersection of demand-responsive transport, intermodal coordination, and multi-agent reinforcement learning. The core difficulty is to keep operations feasible in real time: orders arrive randomly, buses depart on fixed schedules, vessels have energy and capacity limits, and once tasks are committed they cannot be freely revised. Unconstrained DRL exploration can propose infeasible actions. Centralized RL also scales poorly because the joint action space grows exponentially with fleet size, making large-scale training and credit assignment difficult. This work therefore uses a carefully designed state machine and feasibility checks to guarantee valid operations, while MADRL improves efficiency under those rules. This provides a practical bridge between DRL and real-world transport operations.
Contributions.
• A state-machine-driven harbor simulator that captures the end-to-end operational flow, including bus reservations, time windows, capacity, and energy constraints.
• A feasibility-first dispatch interface that bridges operational rules and MADRL: the state machine guarantees valid actions, and learning optimizes efficiency within those rules.
• An empirical demonstration of CTDE MAPPO against a matched heuristic baseline under identical observations and constraints, isolating the benefit of decentralized coordination.

Methodology. A custom event-driven simulator models a 10-dock waterway network with 6 electric vessels, stochastic on-demand arrivals, and deterministic bus transfers. The control problem is formulated as a Decentralized Partially Observable Markov Decision Process (Dec-POMDP) with feasibility-masked actions that enforce time-window, capacity, and energy constraints. A Centralized Training with Decentralized Execution (CTDE) framework based on Multi-Agent Proximal Policy Optimization (MAPPO) is employed, with recurrent (LSTM) policies as the primary configuration. A feedforward MLP variant with matched encoder depth/width is used for architectural ablation.

Results. The learned MADRL policy is evaluated against a value-greedy heuristic baseline under identical observation exposure, constraints, and deterministic execution. Across 40 evaluation episodes with fixed random seeds, the MADRL policy achieves a 14.7% higher episodic reward, 22.0% more completed orders, 44.4% lower deadhead ratio, and 13.5% shorter passenger wait times. Bus-transfer order completion improves by 30.9%, consistent with stronger anticipation of scheduled arrivals. The LSTM–MLP ablation yields comparable deterministic KPIs, indicating that MAPPO performance is robust to the network core when the engineered observation representation is informative.

Conclusion. These findings indicate that intermodal harbor operations can be framed as a multi-agent sequential decision problem with feasible, real-time dispatch under the modeled assumptions. More broadly, the event-driven state-machine design provides a practical bridge between DRL and constrained transport operations, offering a starting point for similar multi-agent, rule-governed systems. The framework and evaluation protocol provide a reproducible basis for future work on larger networks, stochastic disruptions, endogenous passenger choice, and sim-to-real deployment.

... Read more

A combined electric public transport and shared e-micromobility system forms a backbone for sustainable urban mobility, yet electrification introduces interacting uncertainties at every decision level. This dissertation embeds uncertainty modeling and passenger behavior jointly into optimization across strategic, tactical, and operational planning. The resulting models design bus charging infrastructure from sparse data, rebalance shared fleets toward perceived usability, and recommend door-to-door trips that remain executable under uncertainty. ... Read more

This dissertation examines how shared mobility providers can manage demand by aligning operational decisions with user behavior. It highlights behavioral factors that generate systematic and predictable demand responses and embeds them into operational models as demand management levers. By integrating social influence into phased deployment and context effects into choice set design, it shows that high-quality, gradual expansion improves profitability and stabilizes adoption, while behaviorally informed menus strengthen competitive performance. ... Read more

Shared bicycle systems have become an important element of sustainable urban mobility, helping to reduce emissions and congestion while improving first- and last-mile connectivity. However, these systems often suffer from imbalances between bicycle supply and demand, especially during peak hours, reducing user satisfaction and system reliability.

This thesis addresses these challenges by proposing a dynamic, time-adaptive repositioning strategy for continuously operating shared bicycle systems. The strategy integrates station selection, bicycle quantity decisions, and routing. It adapts the planning horizon based on the time of day using shorter routes during peak periods and longer routes during off-peak hours and at night. User satisfaction is captured through a no-service penalty, while operational efficiency is ensured through time-dependent constraints on route duration.
The model is further extended to account for mixed bicycle fleets and applies vehicle-aware repositioning to improve coordination and reduce computational complexity.

Real-world case studies in Zaragoza and Valladolid show that the dynamic strategy significantly outperforms both static and no-repositioning methods. In Zaragoza, it enabled approximately 400 additional satisfied trips per day, while in Valladolid it achieved exceeding a satisfaction level of 95% despite a high demand for electric bicycles. The dynamic approach also proved robust to a 50% increase in demand for both systems, with only minimal reductions in satisfaction levels.

In summary, this thesis demonstrates that a time-adaptive and integrated repositioning strategy can greatly enhance service levels and user satisfaction in shared bicycle systems. The approach supports broader adoption and strengthens the role of shared bicycles in sustainable urban transportation. Future research could build on these findings by targeting bottlenecks with customized strategies, managing broken bicycles, which reduce dock availability, and accounting for solar-powered stations, without a recharge option, where low battery bicycles occupy dock space. ... Read more

Urban public transportation often exhibits pronounced spatio-temporal imbalances in passenger demand, resulting in capacity shortages during peak periods and excessive idle capacity during off-peak times. Meanwhile, with the expansion of e-commerce, urban freight demand continues to grow, making the use of idle public transportation space for freight transport an ideal approach. This study proposes an optimization framework for integrated passenger-freight co-transportation using Modular Autonomous Vehicles (MAVs), formulating a Mixed-Integer Quadratically Constrained Programming (MIQCP) Path-based model based on a space-time network to address the Modular Autonomous Unit (MAU) Routing Problem. The model integrates Fixed-Route Transit (FRT) and Demand-Responsive Transit (DRT), allowing MAUs to dynamically couple/decouple across different routes to meet the spatio-temporal demands of passengers and freight, with the objective of minimizing total operational costs. To tackle the computational complexity of large-scale instances, a customized Adaptive Large Neighborhood Search (ALNS) algorithm is designed, incorporating two initial solution generation methods (GUROBI and Greedy heuristic) and iteratively optimizing solutions through destroy and repair operations. A real-world case study based on the Shanghai bus network validates the effectiveness of the proposed approach. The results demonstrate that the MAU co-transportation system can effectively utilize vehicle compartment space to simultaneously transport passengers and freight, significantly reducing empty load rates, leading to a substantial reduction in operating costs. Without using the co-transportation mode, the number of MAUs used would increase significantly, accompanied by a 3.9% cost increase. Compared to the traditional combination of public transit and delivery vans, costs are reduced by 83.4%. This co-transportation modular transit system, with its unique flexibility, can provide efficient and low-cost transportation services for both passengers and freight within cities. ... Read more

Urban congestion is an increasing challenge for accessibility, travel times and livability. Although on-demand water cabs offer a flexible and sustainable alternative, the design of optimal mooring locations has been underexplored. To address this problem, this article presents a new method for the strategic placement of mooring locations. The proposed method involves identifying potential mooring locations and applying a customised Adaptive Large Neighbourhood Search (ALNS) to manage computational complexity. The model minimises weighted travel time and cost indicators for all origin destination pairs by selecting between public transport and water cab for each trip. The method was evaluated in a case study of Rotterdam. The findings indicate that the proposed ALNS algorithm outperforms both Large Neighbourhood Search (LNS) and Mixed-Integer Linear Programming (MILP) in computation time while simultaneously generating high quality solutions. Moreover, the model is shown to be capable of identifying an efficient network, resulting in greater coverage, shorter travel times and improved accessibility. These results demonstrate that the method constitutes a powerful tool for policymakers in designing robust, efficient and future proof water cab networks. ... Read more

Meal delivery platforms operate in dynamic environments, facing complex challenges including fluctuating customer demand, deciding where and which restaurants to offer, and efficient rider allocation. This thesis introduces the Restaurant Selection and Rider Dimensioning Problem (RSRDP) by proposing an integrated optimization model that jointly determines the optimal restaurant assortments and rider allocations to maximize the platform's expected profitability, while maintaining high service quality standards.

A nested logit model captures customer purchasing behavior on the platform to model how customers choose a restaurant to order from, where customers initially choose a cuisine type and subsequently select a specific restaurant within that type. Rider operations are modeled through a spatial-temporal network, enabling optimized management of rider flows across urban zones. We formulate the problem as a Mixed-Integer Linear Program (MILP) and research advanced solution methods including Benders decomposition, column generation, branch-and-price, and heuristics. Finally, we propose our novel Iterative Assortment Generation (IAG) algorithm to solve the problem, ensuring computational feasibility for large-scale scenarios.

Computational experiments based on simulated urban delivery data illustrate that integrating restaurant selection with rider dimensioning consistently increases profitability and maintains high service quality compared to traditional methods that separate these decisions. Furthermore, we explore the impact of two distinct rider compensation policies: commission-based (payment per delivery) and fixed employment (hourly wages). Our findings reveal significant trade-offs between these compensation structures. Commission-based compensation offers greater operational flexibility, adaptability to fluctuating demands, and cost-effectiveness; however, it also leads to increased rider relocations and potentially reduced rider satisfaction due to uncertainty in workload. Conversely, fixed employment provides stable rider availability, promoting better workforce management, but potentially incurs higher operational costs and decreased responsiveness to short-term demand fluctuations.

Based on these insights, we recommend meal delivery platforms adopt integrated optimization models as part of their tactical planning to effectively balance service quality and profitability. Platforms should carefully evaluate their rider compensation policies, recognizing that the ideal choice may depend on the specific context of their operational environment, such as demand variability, competitive landscape, and workforce preferences.

Future research could build on this study by investigating dynamic compensation schemes that adapt in real-time to observed demand conditions. Additionally, incorporating hybrid compensation schemes, combining salaried and commission-based riders, and utilizing the strengths of both policies, provides a promising research direction. Developing multi-objective models that balance profitability against service quality measures, such as timely deliveries or rider satisfaction, would further guide operational decisions and customer and rider retention. Furthermore, future research could improve the way service districts are designed. Instead of using fixed boundaries, districts could be adjusted dynamically and tailored more precisely to specific customer segments and changing demand patterns. This approach would mean that restaurant offerings can also change throughout the day to closely match customers' preferences, potentially increasing overall profitability. Furthermore, modeling couriers individually, rather than as aggregated flows, allows more precise scheduling that better accounts for realistic travel times, shift durations, and rider constraints. Although this approach increases modeling complexity and computational demands, it offers more accurate operational insights, leading to better-informed management decisions. Finally, employing heuristic-driven column-generation methods and comparing solution approaches systematically can improve scalability and accuracy. ... Read more

To aid the efforts of electrifying public bus systems and to reduce the impact public transit has on emissions in cities, it is necessary for planners to be well informed when making the decision to electrify the transport system. The methodological approach shows that there is value in considering the development of a new transport system over simply electrifying an existing transport network, as there is a trade-off between investment characteristics and passenger satisfaction. Significant cost savings could be realized by jointly considering line planning, infrastructure location, vehicle, battery and fleet size. This has been investigated by developing a decision making model that selects a lineplan seeking to maximize passenger satisfaction, whilst taking into account, vehicle scheduling, battery capacity and charging station location. The model demonstrated that there is a clear trade-off between these aspects and that a multitude of pareto optimal solutions that can be obtained, depending on weighting that decision-makers assign to the objectives. The simplicity of turning the objective of this methodology into a revenue-maximization function, should make this model a suitable starting point for more serious modelling of investment decision in the future. ... Read more

This research provides insight into ways to increase accessibility for Planned Special Events (PSEs) via discrete choice modelling. A special focus within this study is put into the AFAS AZ Stadium in Alkmaar. During the events held at the AFAS AZ Stadium disturbances are experienced by its visitors with regard to crowding levels and access/egress times. Quantifying the traffic volume on the infrastructure around the venue is done via revealed preference methods. Cyclists are counted via pneumatic tubes. Whereas pedestrians are measured via radio-wave sensors. To find ways to mitigate these problems, interviews are held with organisers of similar PSEs within The Netherlands.

To find the effectiveness of the obtained accessibility measures a stated preference study is held. From this study it is obtained that the transportation mode habit that one has is a key factor within the mode choice. Besides the effect on accessibility additional decision factors are discussed to find the best measure which can be applied best at the AFAS AZ Stadium. Based on the decision factors and the performance of the accessibility measures a total of six effective measures are determined. ... Read more

In the airline industry revenue and inventory is managed tightly, with the intent to sell each seat at the right price to the right customer at the right time. This does not always go to plan and seats remain unsold, or demand can be greater than what is available. The question then becomes how can capacity and revenue be managed more efficiently? How can capacity and revenue be maximized to ensure that seat spoilage is minimized. This thesis takes a look at this problem and aims to solve this problem by dynamically allocating capacity as well as becoming paired with revenue management processes. By being able to move capacity around to different flights, it is believed that revenue can be maximized and that seat spoilage can be minimized, creating a more effective use of a perishable product. The goal of this thesis is to show that increases are possible, even if by small margins. This thesis looks to use real-world data to help support its findings along with a random distribution to pair alongside a Monte Carlo simulation to help ensure accuracy. In this data driven methodological approach revenue management and a capacity swapping algorithm are integrated. The revenue management module creates a series of seat inventory allocations as well as demand forecasting. This information is then fed into a simulation which will then simulate the entire booking life- cycle of a fight. Here each flight will build its passenger load and will swap aircraft with other aircraft of varying capacities when the algorithm has determined it is necessary and would create a positive increase in revenue and passenger counts. To assist in this approach a further random forest model is developed to help make informed decisions. These changes are further evaluated by the algorithm to ensure that every change that was made was beneficial and if so will accept the swap and continue to build loads as this new flight. Different case studies were developed that introduced different timing scenarios as well as different decision making tools. The two different decision making tools would use either the random forest model that was developed or a simple logic model that would look at estimated final revenue and estimated final passenger loads. The timing scenarios would look to see when swapping would be needed, whether it is a true daily swap or if swaps can be performed once a month. Ultimately there were positive results, with each scenario tested yielding positive increases in the three KPIs examined (Load Factor, Passengers, Revenue). Each scenario involved testing either the simple logic, or random forest model with varying swapping intervals. The best performing scenario was one where flights were allowed to be swapped daily using a random forest decision model. In this best scenario load factor saw an increase of 1%, the number of passengers increased 1.4%, and revenue increased 0.9%; all against the baseline scenario where zero flights were swapped. This translates into a 36,000 USD weekly increase for a small sample in a route network. Across a full route network, even greater increases can be achieved. ... Read more

Nowadays, urban areas are exposed to various challenges such as climate change, social inequalities, and congestion. Mobility hubs present the opportunity to reshape our cities and mitigate the previously mentioned problems by contributing to a more sustainable and equitable transport system. This thesis defines mobility hubs as places where shared cars, mopeds, and e-bikes are offered to improve connectivity and ameliorate mobility options. Given a limited budget, cities would like to optimize the locations of mobility hubs to maximize benefits. This problem is solved in this thesis by presenting an optimization model that allows the distribution of mobility hubs and allocation of shared cars, mopeds, and e-bikes to maximize social welfare. The algorithm can provide the optimal locations for the hubs and their respective capacity in terms of vehicles, while accounting for multimodal trips. It focuses on maximizing the utility of the population rather than the operators’ profits. The model is divided into several modules: computational modules that calculate the number of people that would like to use a mobility hub; a mathematical optimization module to optimize the capacity, availability, and relocation of shared vehicles; and finally, a genetic algorithm that performs several iterations to find the optimal distribution of hubs. The model developed is applied in a case study for the city of Amsterdam. Several scenarios are performed to assess how the distribution of hubs varies depending on the budget provided to construct them. The results show that having more hubs with a lower number of shared vehicles is more beneficial than having fewer with more vehicles. Areas with higher population density are prioritized when lower budgets are invested in building the hubs. Additionally, the shift towards shared modes and the travel time savings are minimal. The benefits increase considerably when investments lead to complete coverage of the area by the network of mobility hubs. A modal split of 5% for the shared modes is expected when Amsterdam is fully covered by 288 hubs. From an environmental point of view, only 32 % of the shared trips replace trips previously made by car, leading to a limited CO2 emissions reduction of 1.27%. To conclude, the model developed is one of the first models that optimizes the location and capacity of multimodal hubs to maximize social welfare by considering multimodal trips. It has the ability to quantify the benefits of introducing mobility hubs depending on the investments made. ... Read more