Automatic planning for railway shunting yards in the Netherlands is a challenging problem, as these yards form a critical link between rolling stock circulation and the timetable. Trains must be decoupled, coupled, serviced, and departed within tight time windows, while infrastructure and train drivers are limited. The resulting planning problem is large-scale, highly constrained, and characterized by strong interdependencies between operational and driver decisions. In the current state-of-the-art approach, driver scheduling is handled by a heuristic procedure embedded within a local search algorithm. While efficient, this restricts the decision space and may yield suboptimal plans.

This thesis investigates how integrating driver-scheduling decisions into the main local search affects solution quality and search performance. Several integration mechanisms are presented, with increasing levels of integration of driver scheduling decisions. Computational experiments on realistic instances show that integrating driver decisions into the local search consistently improves performance compared to the state-of-the-art approach. Under small time budgets, the greatest improvements are obtained when only part of the driver decision space is integrated, allowing the search to correct harmful heuristic decisions while keeping the search space manageable. With larger time budgets, more integrated methods catch up, indicating that the richer decision space can be exploited when sufficient time is available.

To manage the increase in search space complexity, a novel two-stage search scheme is introduced in which the algorithm first optimizes while keeping driver decisions heuristically decided, and subsequently activates explicit driver-related decision variables. Experiments show that dividing the available computation time between these two stages yields better results than allocating the entire time budget to either stage. This demonstrates that progressive enlargement of the decision space can effectively balance model expressiveness and computational tractability.

Biomanufacturing involves the large-scale production of bio-based products, for example, food ingredients. In fermentation-based factories, living organisms are used to produce the active components of these bio-based products through fermentation in large bioreactors. Modern biomanufacturing plants are highly complex, with many tanks and interconnected pieces of equipment. They are typically designed to make multiple products (in parallel), each with its own recipe and production process. As a result, it is not possible to repeat the same weekly schedule—customer demand changes over time, and different customer orders require different products and thus production schedules.
A feasible schedule should consider several scheduling rules essential for biomanufacturing. For example, in the food industry, it is crucial that tanks are cleaned before, after, or between specific production operations. The production of a single product involves multiple unit operations, for example, starting from fermentation, followed by several filtration steps. Between the different stages, there are rules about how long an intermediate product can wait at a single tank; otherwise, a product risks expiring. Due to the biological nature of fermentation processes, process durations are uncertain and can vary from batch to batch. Due to this complexity, limited resource capacities, and various sources of uncertainty, it poses a significant challenge for factory managers and production planners to fulfill all customer orders on time.....

In recent years, artificial intelligence has increasingly permeated society, including applications in high-stakes domains that may have a significant and even harmful impact on people. For these domains, it is essential to have comprehensible and reliable models. However, many of the most commonly used models, such as neural networks, are black-box models, i.e., inherently opaque and thus hard to trust.

Therefore, this dissertation explores decision trees as a human-interpretable alternative. Specifically, we focus on optimal decision trees, since the traditional greedy heuristics often learn decision trees that are too large to interpret easily, whereas optimal decision trees provably optimize the objective of a learning problem (e.g., accuracy) given a constraint on model size. The compactness and transparency of these optimal models contribute to their interpretability, while in many cases they achieve accuracy comparable to that of much more complex black-box models. In addition, with many optimal learning methods it is easier to specify a learning task other than the default of maximizing classification accuracy.

However, since learning optimal decision trees is an NP-hard problem, scalability is a major concern. State-of-the-art model-based approaches, such as mixed-integer programming and Boolean satisfiability, often do not scale beyond a few hundred samples and shallow depth limits. On the other hand, search-based approaches, such as dynamic programming, scale much better by exploiting the separable tree structure using specialized algorithms to solve subtrees as independent (and repeated) subproblems. But these methods are mostly limited to classification, while their scalability depends on a coarse binarization of the numerical data.

Therefore, in this thesis, we study optimal decision trees across four challenges: (i) efficiently optimizing learning tasks other than classification, (ii) improving our understanding of the differences between optimal and traditional greedy decision trees, (iii) efficiently handling data with numerical attributes; and (iv) efficiently finding the whole Rashomon set: the set of all (near-)optimal decision trees.

In the first challenge we address improving scalability for optimizing decision trees for a variety of learning tasks other than classification. We build on existing dynamic programming approaches for their scalability, and because this scalability depends on the property that subproblems are separable, we study and formalize the precise necessary and sufficient conditions for separability. These conditions turn out to be broader than previously assumed, and we use these new insights to improve the scalability of optimal decision trees for a variety of learning tasks, such as classification under fairness constraints, survival analysis, regression, prescriptive policy learning, and cost-sensitive classification. We also show how different optimization techniques, such as branch-and-bound and a specialized subroutine for depth-two trees, can be adapted and applied across several of these tasks. Together, these techniques improve scalability by one or more orders of magnitude over the state of the art for multiple learning tasks.

The second challenge is to improve our understanding of the differences between optimal and greedy heuristic methods for learning decision trees because previous comparisons were severely limited by lack of scalability. We empirically study the effect of the loss function and hyperparameter tuning of optimal decision trees and conclude that the heuristic and optimal approach are inherently different, for example, concave loss functions, which are required for many heuristics, do not work well for optimal methods. Previous comparisons also resulted in divergent and contradicting claims about optimal methods, and therefore, we empirically test six such claims from the literature. We confirm the most important one: when measuring both size (as a proxy of interpretability) and accuracy, optimal decision trees clearly dominate trees learned by greedy heuristics. We also refute several claims, including the claim that optimal methods are more prone to overfitting, and the claim that the differences between the two methods diminish with more data. For both claims, we find evidence suggesting the opposite.

The third challenge is to improve scalability when the data contains numerical attributes, without falling back to a coarse binarization as most dynamic programming approaches do. Therefore, we develop new algorithmic techniques that exploit the properties of numerical attributes. We show how reasoning about subproblem similarity makes it possible to prove that large parts of the search space can be pruned, while the ordering among numeric values makes it possible to compute aspects of the problem incrementally. Together, these techniques result in scalability improvements of one to two orders of magnitude compared to the state of the art.

For the fourth challenge, we go beyond learning a single optimal model. Instead, we focus on the set of all decision trees within a small percentage of the optimal solution, the so-called Rashomon set. The Rashomon set can be useful for data analysis, for example in determining which attributes have the greatest impact on predictions, or in finding better counterfactual explanations. Our contribution is a more efficient, anytime, sorted enumeration of all decision trees in the Rashomon set using lazy evaluation, as well as adapting and applying existing algorithmic techniques for optimal decision trees to compute the Rashomon set. Here too, our method improves scalability by one or more orders of magnitude compared to the previous best approach.

In summary, this dissertation improves scalability by several orders of magnitude, broadens the range of possible applications, and deepens both the theoretical and empirical understanding of optimal decision trees and decision tree Rashomon sets. Therefore, we can now recommend the use of optimal decision trees as human interpretable models for a large variety of real-world data science problems.

During the daily operation of the railway network, ProRail is responsible for handling delays and planning ad hoc train movements. Train handling documents aid the traffic controllers in common situations. But when multiple trains are delayed, and these documents do not apply, they are left to their own expertise.
In this thesis, we introduce FlexSIPP, an algorithm to plan or replan agents in an existing multi-agent plan. FlexSIPP builds upon the prior works of any-start-time safe interval path planning, where the current routes of the agents are seen as moving obstacles. FlexSIPP loosens this restriction by introducing flexibility: the ability for an agent to delay its plan while minimally impacting other agents.
This algorithm is evaluated on the Dutch railway network. By finding tipping points, that is, the moment it is better to switch the order of two trains on the track to minimize the delay, we can recreate train handling documents. We show that FlexSIPP finds the same solutions within a minute in the case that no other trains are delayed. This implies that FlexSIPP is also able to aid traffic controllers in the case that other trains are delayed

Optimization models are widely used in energy system planning to identify cost-effective investment strategies. However, relying solely on a single optimal solution can be misleading, as it fails to account for model uncertainty, competing objectives, and stakeholder preferences. To address this, near-optimal alternatives, solutions that are close in cost to the optimum but structurally different, are increasingly used to support robust and flexible decision-making.

This thesis explores the generation and evaluation of near-optimal alternatives within energy systems, with a focus on improving the decision relevance of the generated alternatives. This thesis introduces a unified analytical framework, formalizing existing Modeling to Generate Alternatives (MGA) methods using weight vector formulations. This formulation enables a clearer comparison of different techniques that generate these alternatives. This analysis highlights the limitations of current evaluation metrics, particularly their inability to distinguish decision-relevant alternatives from decision-irrelevant ones.

To overcome this gap, the thesis proposes a novel evaluation metric based on dominance relations from multi-objective optimization. This metric identifies non-dominated alternatives, those not strictly worse than any other across all decision variables, as decision-relevant. The thesis introduces a new method that uses Directionally Weighted Variables to generate alternatives aligned with this dominance criterion.

The proposed approach is evaluated using a stylized energy investment model and benchmarked against existing MGA techniques. Results show that traditional methods tend to generate fewer non-dominated alternatives, while the new method generates more non-dominated alternatives within the near-optimal space. This work contributes a new perspective on alternative generation, bridging the gap between mathematical optimality and practical decision support.

As power systems increasingly rely on renewable energy, grid services traditionally supplied by central plants must increasingly be sourced from distributed energy resources (DERs). Virtual power plants (VPPs) aggregate DERs to act as a single entity, but coordination is complicated by information asymmetry, possibly resulting in strategic behaviour. This thesis studies how we can design a mechanism for a commercial VPP, having to satisfy a fixed commitment, while optimising the revenue from the VPP operator.
We first develop a tractable, multi-period VPP model with linear costs, local and temporal constraints for DERs and soft system-wide commitments enforced via deviation penalties. On top of this model we design and compare four mechanisms: first-price sealed bid (FPSB), uniform pricing, Vickrey–Clarke–Groves (VCG) and Arrow–d’Aspremont– Gerard-Varet (AGV). We evaluate them on revenue optimality, weak budget balance, incentive compatibility, individual rationality and scalability. Furthermore, we investigate how the composition of a VPP’s portfolio could inform mechanism design choices. FPSB realises payments equalling costs under truthful reports and remains competitive for small strategic fractions, but overpayment grows with the share of strategic agents and with cost dispersion. Uniform pricing is comparatively insensitive to the strategic fraction but highly sensitive to cost dispersion, often leading to large overpayments. VCG is strategy-proof and insensitive to strategic behaviour, yet externality payments increase with cost dispersion and raise total payouts. AGV keeps the payment-to-cost ratio near or below one by relying on expected externalities and scaling, improving operator viability but potentially violating individual rationality in instances. These results yielded the following guidelines regarding the suitability of mechanisms. FPSB for low strategic participation, uniform pricing for homogeneous portfolios, VCG when truthfulness is vital and external funding is possible, and AGV when operator viability is the hard constraint with safeguards for individual rationality.F

Decision diagrams have steadily become more prominent in the field of combinatorial optimization, being able to outperform the state-of-the-art in e.g. scheduling problems[13]. They have proven even more capable with the introduction of methods such as decision diagram-based Branch & Bound. Often, layer-based binary decision diagram (BDD) encodings are used to encode the problem domain, where a layered structure determines the decisions and decision variables. Recently, state-based multivalued decision diagram (MDD) encodings have also been considered. These do not rely on a layered structure, as instead, each state makes its decisions on the variables independently of other states. This allows for more flexible decision-making as states do not have to compromise with other states on the decisions they make.

This thesis compares these newer state-based MDDs with the commonly used layer-based BDDs, while also introducing and evaluating heuristics for the state-based MDDs. These heuristics include a new beam restriction heuristic that limits the branching factor of the MDDs, a dynamic variable ordering strategy adapted for the new state-based context, and a new local bound for the maximum independent set problem (MISP).

The experiments of this thesis show that the state-based MDDs generally outperform the layer-based BDDs in both runtime and search tree size. Only for some instances with low graph densities, the state-based MDDs were slower. This is resolved with the newly introduced beam restrictions, which can significantly lower the runtime. This speedup is also shown for the graph coloring problem, although the application of the beam restrictions is not as straightforward there as with the MISP. Both the dynamic variable ordering and the new local bound for the MISP show great promise in increasing the efficiency of the search, but are both held back by the additional overhead they introduce. Fortunately, these two techniques can share the added overhead while gaining the combined benefits, resulting in great performance for the MISP when both methods are used together.

This study investigates the performance of three dynamic scheduling approaches—proactive, reactive, and STNU-based—for solving the Multi-Mode Resource-Constrained Project Scheduling Problem with maximal time-lags and no-wait constraints (MMRCPSP/max) in uncertain environments. The performance of the approaches is validated on three key performance measures: solution quality (makespan), offline computation time, and online computation time. Drawing from the current work on stochastic RCPSP/max, this research introduces a more realistic formulation of the problem with multi-mode execution, generalised time lags, and no-wait constraints, under the PyJobShop library and stochastic duration modelling. Experimental results, based on altered PSPLIB instances, show that the three algorithms yield similar feasibility rates. However, they show distinct trade-offs in terms of solution quality, time offline, and time online. The proactive algorithm requires the shortest offline and online computational times, but provides slightly worse makespans. The STNU-based algorithm produces the best average solution quality, with significantly higher offline time. The reactive algorithm offers competitive offline times and solution quality, but has the largest online computational time. The findings provide insight into the trade-offs in solution quality and computational time in dynamic project scheduling and offer insight for using the strategies in the real world.

Modern manufacturing systems must meet hard delivery deadlines while coping with stochastic task durations caused by process noise, equipment variability, and human intervention. Traditional deterministic schedules break down when reality deviates from nominal plans, triggering costly last-minute repairs. This thesis combines offline constraint-programming (CP) optimisation with online temporal-network execution to create schedules that remain feasible under worst-case uncertainty.

First, we build a CP model of the flexible job-shop with per-job deadline tasks and insert an optimal buffer Δ* to obtain a fully pro-active baseline. We then translate the resulting plan into a Simple Temporal Network with Uncertainty (STNU) and verify dynamic controllability, which guarantees that a real-time dispatcher can retime activities for every bounded duration realisation without violating resource or deadline constraints. Extensive Monte-Carlo simulations on the open Kacem 1–4 benchmark suite show that our hybrid approach eliminates 100 % of deadline violations observed in state-of-the-art meta-heuristic schedules, while adding only 3 – 5 % makespan overhead. Scalability experiments confirm that CP solve-times and STNU checks remain sub-second on medium-size instances.

The work demonstrates how temporal-network reasoning can bridge the gap between proactive buffering and dynamic robustness, moving industry a step closer to truly digital, self-correcting factories.

This report investigates the effectiveness of Simple Temporal Networks with Uncer- tainty (STNUs) for solving the Stochastic Flexible Job-Shop Scheduling Problem with Sequence-Dependent Setup Times (SFJSP-SDST), comparing it against proactive and reactive Constraint Programming (CP) approaches. Using a benchmark dataset with varying noise levels, the study evaluates solution quality, feasibility, and computational cost. Results show that the reactive method achieves the lowest makespan due to its dy- namic rescheduling capability but incurs high online computation time. The proactive method offers fast execution, while the STNU-based approach provides a dynamically controllable schedule, albeit with conservative makespans.

This study investigates scheduling strategies for the stochastic duration flexible job-shop problem with no-wait and general time lags constraints (FJSP/NW-GTL). Progress in Constraint Programming (CP) and temporal-networks has renewed interest in assessing the strengths and limitations of different proactive and reactive scheduling approaches. This paper covers the application of a CP-based fully proactive method, a reactive method and STNU-based method on the FJSP/NW-GTL problem comparing results in terms of predetermined objectives and feasibility. In addition, the paper aims to answer how different distributions for task duration affect feasibility and performance. Our results show that strictly proactive methods are infeasible for no-wait constraints and very tight schedules, which lead to adding an online step in the proactive implementation to pursue the comparison between the approaches. With this change, plotting the average makespan across methods by distribution shows that there is not much fluctuation between distribution types in terms of makespan. Moreover, it appears that the proactive method performs the best, followed closely by the reactive method, while the STNU approach results in a notably higher makespan for the same instances. Notably, in terms of feasibility, the proactive and reactive approach have 100% rate of success compared to the STNU approach which is infeasible on 35% of the instances in the dataset.

The transition towards renewable energy requires long-term energy system planning, which depends on solving constrained optimization (CO) problems. These CO problems are becoming increasingly complex, particularly due to the variability introduced by renewable energy sources. Traditional optimization methods struggle to scale with this growing model complexity. In contrast, machine learning approaches allow faster solution computation, but offer only approximate solutions with no guarantees on solution quality, thereby limiting reliability and interpretability in planning applications.

This research addresses these limitations by exploring the use of neural networks to predict feasible solution pairs for the primal and dual formulations of the CO problem, enabling approximate solutions accompanied by bounds on suboptimality. Self-supervised primal-dual learning (PDL) is adapted and extended to produce paired feasible solutions for both the primal and dual formulations of an economic dispatch problem. Feasibility in the primal network is enforced through differentiable repair and completion layers, including novel domain-specific extensions that adjust supply-demand balancing priorities, which were found to be essential for predictive accuracy. This then exposes a structural limitation of PDL: while repair and completion layers are essential for primal learning, they prevent the learning of meaningful dual predictions. To address this, the dual network is trained using a modified loss function that directly optimizes the dual problem, enabling the use of a completion layer adopted from the literature to ensure dual feasibility. An additional novel classification-based layer that incorporates prior knowledge on the dual variables further improves the dual prediction quality when applied to the economic dispatch problem. Finally, the trained networks are integrated into Benders decomposition, a technique that breaks CO problems into easier, independent problems. This enables a hybrid approach: approximate solutions with bounds on suboptimality can be obtained swiftly, whereas exact solving remains available if necessary, retaining all theoretical convergence and optimality guarantees and potentially reducing computational time.

The proposed framework is evaluated on a generation expansion planning problem, where the economic dispatch subproblems are solved iteratively using the trained networks. The results demonstrate a theoretically grounded and empirically validated proof of concept for producing solutions with quality guarantees using learning-based methods, which is generally applicable to any problem compatible with Benders decomposition where the resulting subproblem admits a conic formulation. However, the results do not show conclusive speed-ups due to a mismatch between the training data and the data encountered during Benders iterations. By demonstrating how neural networks can be used to generate approximate, feasible solutions accompanied by theoretical guarantees on solution quality, this research contributes to the advancement of scalable and reliable constrained optimization methods for energy systems.

Production planning in the biomanufacturing sector presents significant challenges due to uncertainties in job durations caused by biological variability, environmental conditions, and raw material quality. Traditional scheduling methods typically fail to adapt to these uncertainties, leading to suboptimal outcomes. This research addresses this issue at DSM-Firmenich, focusing on optimizing production planning while maximizing profit, adhering to deadlines, and efficiently utilizing resources. We propose an integrated approach using Mixed-Integer Linear Programming (MILP) and Constraint Programming (CP) models, alongside Probabilistic Simple Temporal Networks (PSTNs) to handle uncertainty in real-time scheduling. The study introduces an offline optimization procedure for proactive scheduling decisions and a reactive real-time algorithm for adjustments of the planned schedule. This work showcases the potential of applying PSTNs in biomanufacturing and sets the stage for future research aimed at enhancing real-time execution strategies in factory environments.

This thesis investigates improving the group project matching algorithm of TU Delft's Project Forum platform. We formalize the matching problem as a many-to-one, one-sided matching with group formation, where students have preferences over project topics and may wish to pregroup with peers.

We implement and evaluate two mechanisms: Chiarandini – a mechanism reimplemented from the literature and adapted to our grouping setting – and our novel Fair mechanism which incorporates additional fairness considerations for pregrouping. Both use mixed integer linear programming. Leximin is selected as the objective function. These two mechanisms are compared to our previous mechanism, BEPSys, which matches greedily determined groups of students to projects using an optimal algorithm.

Our evaluation uses historical instances from TU Delft and SDU as well as synthetic instances that we generate. Results show that both new mechanisms significantly outperform the prior BEPSys algorithm, both in terms of quality of results and runtime. Our novel Fair mechanism successfully allows pregrouping without disadvantaging solo students, although sometimes at significant cost to pregrouping students. The leximin objective, implemented using an ordered weighted averaging function, is shown to not work optimally in some tested instances.

Decision-Focused Learning (DFL) focuses on a setting where a system gets as input some features and needs to predict coefficients to a downstream optimization problem. Classically, one would apply a two-stage solution, which trains the predictor as a regression task and only uses the optimizer during evaluation. However, the two-stage solution fails to optimize the downstream optimization problem. As such, one might use DFL techniques to train the predictor. Nonetheless, these fail to take the entire solution space into account and only optimize toward the optimal true solution, and as such, they might fail to optimize the total downstream value. Furthermore, these techniques are computationally expensive.

We tackle these two problems using Decision Diagrams (DDs) for DFL. DDs for DFL have three main benefits. First, the DD can be cached between runs, speeding up the training loop. Second, we present four novel loss functions that use DDs to reason efficiently over entire solution spaces. Furthermore, we introduce a novel method to relax the DDs to reduce solve-time during training.

We experimentally show that the DDs speed up the training loop substantially. We further show that the DFL losses perform on par with other state-of-the-art DFL losses. Finally, we experimentally show when and which losses work with relaxed DDs.

The research in this thesis falls within the realm of optimization under uncertainty, a crucial area in computer science and mathematics with broad applications in power systems, finance, machine learning, healthcare, and more. This thesis presents three main contributions across electric vehicle charging scheduling, decision-focused learning, and reinforcement learning. Beyond advancing the state of the art in each of these domains, our contributions emphasize the importance of effective problem formulations

Artificial intelligence (AI) has become a widely discussed and transformative technology, with its adoption growing across industries to drive insights and impact. In this thesis, we explore how AI methods and algorithms can facilitate the operation of soft-fruit supply chains, using strawberries as a case study.

The thesis begins by presenting the general background and various perspectives from related works on how AI and machine learning (ML) have been applied to address problems in agricultural or horticultural practices. This includes tasks that, while not directly optimizing supply strategies, still contribute to solving broader challenges. In a nutshell, this thesis categorizes the scope of study into three scales: the single-fruit scale, the greenhouse scale, and the market scale. Within each scale, we review the existing research, identify knowledge gaps, and introduce robust and applicable methodologies capable of dealing with real-world conditions.

Since no publicly available datasets met the requirements of the research plan, we established several datasets for research on the soft-fruit supply chain through collecting, annotating, and (pre-)processing data. These newly curated datasets not only support the research presented in this thesis but also lay a foundation for future research from various perspectives. Details about these datasets are introduced in Chapter 2. Moreover, we conceptualize the process of gathering longitudinal observations from growth monitoring images as a multiple object tracking (MOT) task. We named the image collection and their MOT annotations as ``The Growing Strawberries (GSD)''. The computer vision challenge that GSD brings are further benchmarked and discussed in Chapter 3. Following this, the core contributions of the thesis is presented from Chapter 3 to 6, each corresponding to a published paper or one currently under review. Finally, Chapter 7 summarizes the research findings, answering the research questions proposed in Chapter 1 and discussing the overall work of the thesis.

We discuss these contributions for each of the three mentioned scales separately:

At the fruit scale, we designed and analyzed novel methodologies to keep track of the fruit growths and to predict key properties, including both external characteristics like ripeness and internal qualities such as sweetness. For the ripeness, we propose to use appearance properties, mainly the hue, as an objective metric to quantify it. For the sweetness, we trained deep neural networks to perform non-destructive prediction using environmental and image data, individually and integrally.

Our employment of color analysis and ML models provides a non-destructive and generalizable manner that ensures consistency when upstream and downstream parties in a supply chain estimate the properties of fruits. Meanwhile, the models perform comparatively with laboratory benchmarks even under imperfect, outdoor data collection. We further demonstrated the model in a mobile app to further facilitate adoption in the field.

By benchmarking state-of-the-art MOT algorithms on GSD, we illustrated the new challenges that are brought by this use case: first, the MOT objects change appearance during the tracking due to their biological development, and second, sparse frame rates introduce irregular movements from image to image. We showcased how fruit properties, such as ripeness, change over its life cycle. The results not only provide quantitative measurements that describe the fruit's biological development, but also depict the pain points of current MOT algorithms' predictions. In the meantime, by quantifying these changes over the biological development, we also retrieval relevant information and datasets to support predictions of the changes.

At the greenhouse scale, we designed a framework that optimizes the timing of fruit harvesting by integrating the aforementioned quantified changes over biological development, based on sequential demands about the desired quantities to be harvested. Essentially, the framework makes fruit-specific decisions on dates of harvests by leveraging the monitoring data. The decisions are thus made to enhance both current and future demand-fulfillment capabilities. At each stage of this framework, we evaluated various methods and discussed their effectiveness in achieving the stage targets. For example, how to process the infield data to achieve coherent functions about the ripeness development, how to predict future changes, how to include different perspectives in the optimization model, and etc. As the decisions are made for each specific fruit, the work also demonstrates significant potential for integration with mobile apps and harvesting robots. On top of that, the information retrieval function can also serve as a standalone application to provide objective fruit-level quality assessment.

At the market scale, we focus on the portfolio optimization of a grower under a widely applied mechanism of the market system: the majority of demands for harvests are predetermined through advance contracts, which also serves as an a priori condition of the solution proposed at the greenhouse level. The local market, with dynamic prices and demands, can be used to save losses from the difference in contracted demands and the actual yield. To mitigate outlying decision failures, we introduced the ``smart predict-then-optimize (SPO)'' method, which trains models to predict future yield and local market prices.
Our results illustrate that SPO loss primarily affects the bias layer in neural networks, contrasting with models trained using mean squared error (MSE). This difference essentially leads to more conservative estimations in decision-making scenarios, and also motivates and highlights the importance of effective MSE-based pre-training. Additionally, our study reveals how SPO loss makes models interact when multiple neural networks are trained to predict decision parameters with diverse functions. This insight expands the applicability of SPO loss across a broader range of use cases and model architectures, underscoring its contribution to the field of decision-focused learning.

In conclusion, this thesis introduces diverse data-driven methodologies to tackle the distinct tasks involved in optimizing fruit supply, using strawberries as a case study. Central to our approach is the effective utilization of data, which serves as the foundation for solutions that span from fruit-level evaluations to market-level planning. By leveraging analytics of non-destructive data, our solutions provide objective estimations of fruit quality, fostering a more consistent shared understanding between sellers and buyers while reducing potential food waste. Overall, these advancements push the boundaries of AI in supporting decision-making during the supply of soft fruits, particularly for smaller growers. The findings not only empower more efficient and sustainable supply chain operations but also highlight the strong potential for many practical real-world applications.

Path finding is an important component in solving a wide array of engineering problems, ranging from video games to real-life applications such as automated warehouse management and autonomous vehicles.
Path finding algorithms are designed to solve complex problems, and in order to do so, assumptions are necessary to simplify the problems.
While these assumptions are important, using them makes the obtained algorithms less applicable to a real-life scenario, and as such, verifying how lifting some of them would affect the obtained results is worth pursuing.

Two main assumptions were identified and subsequently lifted.
First, classic multi-agent path finding algorithms use a centralized approach, where solutions are computed before execution.
This results in a long computation period followed by execution. Lifting this assumption results in a decentralized approach where agents solve conflicts on the go, while approaching their target.
The second assumption made by state of the art algorithms is that agents participating in a multi-agent path finding problem share a common goal: minimizing a global cost function.
This is not always applicable, as in a real-life scenario participating agents can have selfish goals.
This assumptions has been lifted by allowing agents to negotiate their paths by trading with the other participants to create better solutions for themselves.

The Selfish Localized Pathfinding (SLP) algorithm has been designed to lift these assumptions. It describes a decentralized algorithm that allows participating agents to negotiate their paths, which makes a good candidate for an application closer to real-life.

The SLP algorithm has been tested in order to evaluate its performance, both in terms of its ability to solve a set of test cases, and in terms of the cost incurred by the participating agents.
SLP performed well in varied domains.
SLP solved significantly more cases than Conflict Based Search, a centralized state of the art path finding algorithm.
This comes at the expense of an increase in the path lengths obtained by the algorithm.
This downside is offset by the significant decrease in the time required to solve problems, which can be divided in small clusters due to the decentralized approach of the algorithm.
On the whole, SLP can provide a good alternative to Conflict Based Search.

When addressing combinatorial optimization problems, the focus is predominantly on their computational complexity, and it is often forgotten to look at the bigger picture. As a result, it is common to miss critical details which could play a major role in the overall process. One such detail is the presence of uncertainty in the real world. A naive approach might directly predict values for the uncertain parameters, without taking into account that the ultimate goal is to obtain sound decisions. Consequently, in many cases, the resulting solutions are suboptimal. This challenge is precisely the premise behind Decision-Focused Learning (DFL) framework, which is a core of this work. This study pioneers the application of the DFL framework to scheduling problems with uncertain processing times, utilizing contextual features to predict these uncertainties. By employing the promising Score Function Gradient Estimation method, the research tackles the issue of non-differentiable regret loss functions in DFL. Key contributions include the development of techniques to enhance the performance of the Score Function, an in-depth analysis of DFL's applicability to complex scheduling scenarios, and a detailed evaluation of its strengths and weaknesses. This work not only demonstrates the potential of DFL in this new context but also lays the groundwork for future research and improvements in handling uncertainty in combinatorial optimization.

Shunting yards are locations next to train stations that serve as parking places for trains when they are not in operation and often contain facilities for maintenance and cleaning for passenger trains. Planning of the tasks regarding shunting trains involves routing, assignment of tracks, and scheduling tasks. This is done manually and requires a lot of effort, making it inefficient. Identifying patterns specific in the arrival times of trains at shunting yards can help to predict future train arrivals and potential delays throughout the year more accurately. This enables the alignment of staff and equipment with train arrivals, minimizing idle time and optimizing cost efficiency.
This research focuses on extracting and analyzing the arrival times of trains at shunting yards using a dataset consisting of GPS data. It conducts two algorithms to cluster the given data for each train unit within and across days to identify the same train across different days. Distributions and heatmaps of the arrival times and delays are created based on the identified train series. They are analyzed to identify patterns in train arrival times and delays across different months.