[pt] Métodos ensemble como random forest, boosting e bagging foram extensivamente estudados e provaram ter uma acurácia melhor do que usar apenas um preditor. Entretanto, a desvantagem é que os modelos obtidos utilizando esses métodos podem ser muito mais difíceis de serem interpretados do que por exemplo, uma árvore de decisão. Neste trabalho, nós abordamos o problema de construir uma árvore de decisão que aproximadamente reproduza um conjunto de árvores, explorando o tradeoff entre acurácia e interpretabilidade, que pode ser alcançado quando a reprodução exata do conjunto de árvores é relaxada. Primeiramente, nós formalizamos o problem de obter uma árvore de decisão de uma determinada profundidade que seja a mais aderente ao conjunto de árvores e propomos um algoritmo de programação dinâmica para resolver esse problema. Nós também provamos que a árvore de decisão obtida por esse procedimento satisfaz garantias de generalização relacionadas a generalização do modelo original de conjuntos de árvores, um elemento crucial para a efetividade dessa árvore de decisão em prática. Visto que a complexidade computacional do algoritmo de programação dinâmica é exponencial no número de features, nós propomos duas heurísticas para gerar árvores de uma determinada profundidade com boa aderência em relação ao conjunto de árvores. Por fim, nós conduzimos experimentos computacionais para avaliar os algoritmos propostos. Quando utilizados classificadores mais interpretáveis, os resultados indicam que em diversas situações a perda em acurácia é pequena ou inexistente: restrigindo a árvores de decisão de profundidade 6, nossos algoritmos produzem árvores que em média possuem acurácias que estão a 1 por cento (considerando o algoritmo de programção dinâmica) ou 2 por cento (considerando os algoritmos heurísticos) do conjunto original de árvores. / [en] Ensemble methods in machine learning such as random forest, boosting, and bagging have been thoroughly studied and proven to have better accuracy than using a single predictor. However, their drawback is that they give models that can be much harder to interpret than those given by, for example, decision trees. In this work, we approach in a principled way the problem of constructing a decision tree that approximately reproduces a tree ensemble, exploring the tradeoff between accuracy and interpretability that can be obtained once exact reproduction is relaxed. First, we formally define the problem of obtaining the decision tree of a given depth that is most adherent to a tree ensemble and give a Dynamic Programming algorithm for solving this problem. We also prove that the decision trees obtained by this procedure satisfy generalization guarantees related to the generalization of the original tree ensembles, a crucial element for their effectiveness in practice. Since the computational complexity of the Dynamic Programming algorithm is exponential in the number of features, we also design heuristics to compute trees of a given depth with good adherence to a tree ensemble. Finally, we conduct a comprehensive computational evaluation of the algorithms proposed. The results indicate that in many situations, there is little or no loss in accuracy in working more interpretable classifiers: even restricting to only depth-6 decision trees, our algorithms produce trees with average accuracies that are within 1 percent (for the Dynamic Programming algorithm) or 2 percent (heuristics) of the original random forest.

Streaming services provide different settings where customers can choose a sound and video quality based on personal preference. The majority of users never make an active choice; instead, they get a default quality setting which is chosen automatically for them based on some parameters, like internet connection quality. This thesis explores personalising the default audio setting, intending to improve the user experience. It achieves this by leveraging machine learning trained on the fraction of users that have made active choices in changing the quality setting. The assumption that user similarity in users who make an active choice can be leveraged to impact user experience was the idea behind this thesis work. It was issued to study which type of data from different categories: demographic, product and consumption is most predictive of a user's taste in sound quality. A case study was conducted to achieve the goals for this thesis. Five predictive model prototypes were trained, evaluated, compared and analysed using two different algorithms: XGBoost and Logistic Regression, and targeting two regions: Sweden and Brazil. Feature importance analysis was conducted using SHapley Additive exPlanations(SHAP), a unified framework for interpreting predictions with a game theoretic approach, and by measuring coefficient weights to determine the most predictive features. Besides exploring the feature impact, the thesis also answers how reasonable it is to generalise these models to non-selecting users by performing hypothesis testing. The project also covered bias analysis between users with and without active quality settings and how that affects the models. The models with XGBoost had higher performance. The results showed that demographic and product data had a higher impact on model predictions in both regions. Although, different regions did not have the same data features as most predictive, so there were differences observed in feature importance between regions and also between platforms. The results of hypothesis testing did not indicate a valid reason to consider the models to work for non-selective users. However, the method is negatively affected by other factors such as small changes in big datasets that impact the statistical significance. Data bias in some data features was found, which indicated a correlation but not the causation behind the patterns. The results of this thesis additionally show how machine learning can improve user experience in regards to default sound quality settings, by leveraging models on user similarity in users who have changed the sound quality to the most suitable for them. / Streamingtjänster erbjuder olika inställningar där kunderna kan välja ljud- och videokvalitet baserat på personliga preferenser. Majoriteten av användarna gör aldrig ett aktivt val; de tilldelas istället en standardkvalitetsinställning som väljs automatiskt baserat på vissa parametrar, som internetanslutningskvalitet. Denna avhandling undersöker anpassning av standardljudinställningen, med avsikt att förbättra användarupplevelsen. Detta uppnås genom att tillämpa maskininlärning på den andel användare som har aktivt ändrat kvalitetsinställningen. Antagandet att användarlikhet hos användare som gör ett aktivt val kan utnyttjas för att påverka användarupplevelsen var tanken bakom detta examensarbete. Det utfärdades för att studera vilken typ av data från olika kategorier: demografi, produkt och konsumtion är mest förutsägande för användarens smak i ljudkvalitet. En fallstudie genomfördes för att uppnå målen för denna avhandling. Fem prediktiva modellprototyper tränades, utvärderades, jämfördes och analyserades med två olika algoritmer: XGBoost och Logistisk Regression, och inriktade på två regioner: Sverige och Brasilien. Analys av funktionsvikt genomfördes med SHapley Additive exPlanations (SHAP), en enhetlig ram för att tolka förutsägelser med en spelteoretisk metod, och genom att mäta koefficientvikter för att bestämma de mest prediktiva funktionerna. Förutom att utforska funktionens påverkan, svarar avhandlingen också på hur rimligt det är att generalisera dessa modeller för icke-selektiva användare genom att utföra hypotesprövning. Projektet omfattade också biasanalys mellan användare med och utan aktiva kvalitetsinställningar och hur det påverkar modellerna. Modellerna med XGBoost hade högre prestanda. Resultaten visade att demografisk data och produktdata hade en högre inverkan på modellförutsägelser i båda regionerna. Däremot hade olika regioner inte samma datafunktioner som mest prediktiva, skillnader observerades i funktionsvikt mellan regioner och även mellan plattformar. Resultaten av hypotesprövningen indikerade inte på vägande anledning för att anse att modellerna skulle fungera för icke-selektiva användare. Däremot har metoden påverkats negativt av andra faktorer som små förändringar i stora datamängder som påverkar den statistiska signifikansen. Data bias hittades i vissa datafunktioner, vilket indikerade en korrelation men inte orsaken bakom mönstren. Resultaten av denna avhandling visar dessutom hur maskininlärning kan förbättra användarupplevelsen när det gäller standardinställningar för ljudkvalitet, genom att utnyttja modeller för användarlikhet hos användare som har ändrat ljudkvaliteten till det mest lämpliga för dem.

State-of-the-art Machine Learning (ML) algorithms show impressive results for a myriad of applications. However, they operate as a sort of a black box: the decisions taken are not human-understandable. There is a need for transparency and interpretability of ML predictions to be wider accepted in society, especially in specific fields such as medicine or finance. Most of the efforts so far have focused on explaining supervised learning. This project aims to use some of these successful explainability algorithms and apply them to Reinforcement Learning (RL). To do so, we explain the actions of a RL agent playing Atari’s Breakout game, using two different explainability algorithms: Shapley Additive Explanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME). We successfully implement both algorithms, which yield credible and insightful explanations of the mechanics of the agent. However, we think the final presentation of the results is sub-optimal for the final user, as it is not intuitive at first sight. / De senaste algoritmerna för maskininlärning (ML) visar imponerande resultat för en mängd olika tillämpningar. De fungerar dock som ett slags ”svart låda”: de beslut som fattas är inte begripliga för människor. Det finns ett behov av öppenhet och tolkningsbarhet för ML-prognoser för att de ska bli mer accepterade i samhället, särskilt inom specifika områden som medicin och ekonomi. De flesta insatser hittills har fokuserat på att förklara övervakad inlärning. Syftet med detta projekt är att använda några av dessa framgångsrika algoritmer för att förklara och tillämpa dem på förstärkning lärande (Reinforcement Learning, RL). För att göra detta förklarar vi handlingarna hos en RL-agent som spelar Ataris Breakout-spel med hjälp av två olika förklaringsalgoritmer: Shapley Additive Explanations (SHAP) och Local Interpretable Model-agnostic Explanations (LIME). Vi genomför framgångsrikt båda algoritmerna, som ger trovärdiga och insiktsfulla förklaringar av agentens mekanik. Vi anser dock att den slutliga presentationen av resultaten inte är optimal för slutanvändaren, eftersom den inte är intuitiv vid första anblicken. / Els algoritmes d’aprenentatge automàtic (Machine Learning, ML) d’última generació mostren resultats impressionants per a moltes aplicacions. Tot i això, funcionen com una mena de caixa negra: les decisions preses no són comprensibles per a l’ésser humà. Per tal que les prediccion preses mitjançant ML siguin més acceptades a la societat, especialment en camps específics com la medicina o les finances, cal transparència i interpretabilitat. La majoria dels esforços que s’han fet fins ara s’han centrat a explicar l’aprenentatge supervisat (supervised learning). Aquest projecte pretén utilitzar alguns d’aquests existosos algoritmes d’explicabilitat i aplicar-los a l’aprenentatge per reforç (Reinforcement Learning, RL). Per fer-ho, expliquem les accions d’un agent de RL que juga al joc Breakout d’Atari utilitzant dos algoritmes diferents: explicacions additives de Shapley (SHAP) i explicacions model-agnòstiques localment interpretables (LIME). Hem implementat amb èxit tots dos algoritmes, que produeixen explicacions creïbles i interessants de la mecànica de l’agent. Tanmateix, creiem que la presentació final dels resultats no és òptima per a l’usuari final, ja que no és intuïtiva a primera vista.

The thesis is dedicated to the approximation of high-dimensional functions from scattered data nodes. Many methods in this area lack the property of interpretability in the context of explainable artificial intelligence. The idea is to address this shortcoming by proposing a new method that is intrinsically designed around interpretability. The multivariate analysis of variance (ANOVA) decomposition is the main tool to achieve this purpose. We study the connection between the ANOVA decomposition and orthonormal bases to obtain a powerful basis representation. Moreover, we focus on functions that are mostly explained by low-order interactions to circumvent the curse of dimensionality in its exponential form. Through the connection with grouped index sets, we can propose a least-squares approximation idea via iterative LSQR. Here, the proposed grouped transformations provide fast algorithms for multiplication with the appearing matrices. Through global sensitivity indices we are then able to analyze the approximation which can be used in improving it further. The method is also well-suited for the approximation of real data sets where the sparsity-of-effects principle ensures a low-dimensional structure. We demonstrate the applicability of the method in multiple numerical experiments with real and synthetic data.:1 Introduction 2 The Classical ANOVA Decomposition 3 Fast Multiplication with Grouped Transformations 4 High-Dimensional Explainable ANOVA Approximation 5 Numerical Experiments with Synthetic Data 6 Numerical Experiments with Real Data 7 Conclusion Bibliography / Die Arbeit widmet sich der Approximation von hoch-dimensionalen Funktionen aus verstreuten Datenpunkten. In diesem Bereich leiden vielen Methoden darunter, dass sie nicht interpretierbar sind, was insbesondere im Kontext von Explainable Artificial Intelligence von großer Wichtigkeit ist. Um dieses Problem zu adressieren, schlagen wir eine neue Methode vor, die um das Konzept von Interpretierbarkeit entwickelt ist. Unser wichtigstes Werkzeug dazu ist die Analysis of Variance (ANOVA) Zerlegung. Wir betrachten insbesondere die Verbindung der ANOVA Zerlegung zu orthonormalen Basen und erhalten eine wichtige Reihendarstellung. Zusätzlich fokussieren wir uns auf Funktionen, die hauptsächlich durch niedrig-dimensionale Variableninteraktionen erklärt werden. Dies hilft uns, den Fluch der Dimensionen in seiner exponentiellen Form zu überwinden. Über die Verbindung zu Grouped Index Sets schlagen wir dann eine kleinste Quadrate Approximation mit dem iterativen LSQR Algorithmus vor. Dabei liefern die vorgeschlagenen Grouped Transformations eine schnelle Multiplikation mit den entsprechenden Matrizen. Unter Zuhilfenahme von globalen Sensitvitätsindizes können wir die Approximation analysieren und weiter verbessern. Die Methode ist zudem gut dafür geeignet, reale Datensätze zu approximieren, wobei das sparsity-of-effects Prinzip sicherstellt, dass wir mit niedrigdimensionalen Strukturen arbeiten. Wir demonstrieren die Anwendbarkeit der Methode in verschiedenen numerischen Experimenten mit realen und synthetischen Daten.:1 Introduction 2 The Classical ANOVA Decomposition 3 Fast Multiplication with Grouped Transformations 4 High-Dimensional Explainable ANOVA Approximation 5 Numerical Experiments with Synthetic Data 6 Numerical Experiments with Real Data 7 Conclusion Bibliography

Fraud is a common crime within the insurance industry, and insurance companies want to quickly identify fraudulent claimants as they often result in higher premiums for honest customers. Due to the digital transformation where the sheer volume and complexity of available data has grown, manual fraud detection is no longer suitable. This work aims to automate the detection of fraudulent claimants and gain practical insights into fraudulent behavior using unsupervised anomaly detection, which, compared to supervised methods, allows for a more cost-efficient and practical application in the insurance industry. To obtain interpretable results and benefit from the temporal dependencies in human behavior, we propose two variations of LSTM based autoencoders to classify sequences of insurance claims. Autoencoders can provide feature importances that give insight into the models' predictions, which is essential when models are put to practice. This approach relies on the assumption that outliers in the data are fraudulent. The models were trained and evaluated on a dataset we engineered using data from a Swedish insurance company, where the few labeled frauds that existed were solely used for validation and testing. Experimental results show state-of-the-art performance, and further evaluation shows that the combination of autoencoders and LSTMs are efficient but have similar performance to the employed baselines. This thesis provides an entry point for interested practitioners to learn key aspects of anomaly detection within fraud detection by thoroughly discussing the subject at hand and the details of our work. / <p>Gjordes digitalt via Zoom. </p>

Machine learning-based systems are rapidly becoming popular because it has been realized that machines are more efficient and effective than humans at performing certain tasks. Although machine learning algorithms are extremely popular, they are also very literal and undeviating. This has led to a huge research surge in the field of interpretability in machine learning to ensure that machine learning models are reliable, fair, and can be held liable for their decision-making process. Moreover, in most real-world problems just making predictions using machine learning algorithms only solves the problem partially. Time series is one of the most popular and important data types because of its dominant presence in the fields of business, economics, and engineering. Despite this, interpretability in time series is still relatively unexplored as compared to tabular, text, and image data. With the growing research in the field of interpretability in machine learning, there is also a pressing need to be able to quantify the quality of explanations produced after interpreting machine learning models. Due to this reason, evaluation of interpretability is extremely important. The evaluation of interpretability for models built on time series seems completely unexplored in research circles. This thesis work focused on achieving and evaluating model agnostic interpretability in a time series forecasting problem.  The use case discussed in this thesis work focused on finding a solution to a problem faced by a digital consultancy company. The digital consultancy wants to take a data-driven approach to understand the effect of various sales related activities in the company on the sales deals closed by the company. The solution involved framing the problem as a time series forecasting problem to predict the sales deals and interpreting the underlying forecasting model. The interpretability was achieved using two novel model agnostic interpretability techniques, Local interpretable model- agnostic explanations (LIME) and Shapley additive explanations (SHAP). The explanations produced after achieving interpretability were evaluated using human evaluation of interpretability. The results of the human evaluation studies clearly indicate that the explanations produced by LIME and SHAP greatly helped lay humans in understanding the predictions made by the machine learning model. The human evaluation study results also indicated that LIME and SHAP explanations were almost equally understandable with LIME performing better but with a very small margin. The work done during this project can easily be extended to any time series forecasting or classification scenario for achieving and evaluating interpretability. Furthermore, this work can offer a very good framework for achieving and evaluating interpretability in any machine learning-based regression or classification problem. / Maskininlärningsbaserade system blir snabbt populära eftersom man har insett att maskiner är effektivare än människor när det gäller att utföra vissa uppgifter. Även om maskininlärningsalgoritmer är extremt populära, är de också mycket bokstavliga. Detta har lett till en enorm forskningsökning inom området tolkbarhet i maskininlärning för att säkerställa att maskininlärningsmodeller är tillförlitliga, rättvisa och kan hållas ansvariga för deras beslutsprocess. Dessutom löser problemet i de flesta verkliga problem bara att göra förutsägelser med maskininlärningsalgoritmer bara delvis. Tidsserier är en av de mest populära och viktiga datatyperna på grund av dess dominerande närvaro inom affärsverksamhet, ekonomi och teknik. Trots detta är tolkningsförmågan i tidsserier fortfarande relativt outforskad jämfört med tabell-, text- och bilddata. Med den växande forskningen inom området tolkbarhet inom maskininlärning finns det också ett stort behov av att kunna kvantifiera kvaliteten på förklaringar som produceras efter tolkning av maskininlärningsmodeller. Av denna anledning är utvärdering av tolkbarhet extremt viktig. Utvärderingen av tolkbarhet för modeller som bygger på tidsserier verkar helt outforskad i forskarkretsar. Detta uppsatsarbete fokuserar på att uppnå och utvärdera agnostisk modelltolkbarhet i ett tidsserieprognosproblem.  Fokus ligger i att hitta lösningen på ett problem som ett digitalt konsultföretag står inför som användningsfall. Det digitala konsultföretaget vill använda en datadriven metod för att förstå effekten av olika försäljningsrelaterade aktiviteter i företaget på de försäljningsavtal som företaget stänger. Lösningen innebar att inrama problemet som ett tidsserieprognosproblem för att förutsäga försäljningsavtalen och tolka den underliggande prognosmodellen. Tolkningsförmågan uppnåddes med hjälp av två nya tekniker för agnostisk tolkbarhet, lokala tolkbara modellagnostiska förklaringar (LIME) och Shapley additiva förklaringar (SHAP). Förklaringarna som producerats efter att ha uppnått tolkbarhet utvärderades med hjälp av mänsklig utvärdering av tolkbarhet. Resultaten av de mänskliga utvärderingsstudierna visar tydligt att de förklaringar som produceras av LIME och SHAP starkt hjälpte människor att förstå förutsägelserna från maskininlärningsmodellen. De mänskliga utvärderingsstudieresultaten visade också att LIME- och SHAP-förklaringar var nästan lika förståeliga med LIME som presterade bättre men med en mycket liten marginal. Arbetet som utförts under detta projekt kan enkelt utvidgas till alla tidsserieprognoser eller klassificeringsscenarier för att uppnå och utvärdera tolkbarhet. Dessutom kan detta arbete erbjuda en mycket bra ram för att uppnå och utvärdera tolkbarhet i alla maskininlärningsbaserade regressions- eller klassificeringsproblem.

Computer simulation is a commonly applied technique for studying electronic warfare duels. This thesis aims to apply machine learning techniques to convert simulation output data into knowledge and insights regarding defensive actions for a ship facing multiple hostile missiles. The analysis may support tactical decision-making, hence the interpretability aspect of predictions is necessary to allow for human evaluation and understanding of impacts from the explanatory variables. The final distance for the threats to the target and the probability of the threats hitting the target was modeled using a multi-layer perceptron model with a multi-task approach, including custom loss functions. The results generated in this study show that the selected methodology is more successful than a baseline using regression models. Modeling the outcome with artificial neural networks results in a black box for decision making. Therefore the concept of interpretable machine learning was applied using a post-hoc approach. Given the learned model, the features considered, and the multiple threats, the feature contributions to the model were interpreted using Kernel SHapley Additive exPlanations (SHAP). The method consists of local linear surrogate models for approximating Shapley values. The analysis primarily showed that an increased seeker activation distance was important, and the increased time for defensive actions improved the outcomes. Further, predicting the final distance to the ship at the beginning of a simulation is important and, in general, a guidance of the actual outcome. The action of firing chaff grenades in the tracking gate also had importance. More chaff grenades influenced the missiles' tracking and provided a preferable outcome from the defended ship's point of view.

[pt] À medida que os modelos de aprendizado de máquina penetram áreas críticas como medicina, sistema de justiça criminal e mercados financeiros, sua opacidade, que impede que as pessoas interpretem a maioria deles, se tornou um problema a ser resolvido. Neste trabalho, apresentamos uma nova taxonomia para classificar qualquer método, abordagem ou estratégia para lidar com o problema da interpretabilidade de modelos de aprendizado de máquina. A taxonomia proposta que preenche uma lacuna existente nas estruturas de taxonomia atuais em relação à percepção subjetiva de diferentes intérpretes sobre um mesmo modelo. Para avaliar a taxonomia proposta, classificamos as contribuições de artigos científicos relevantes da área. / [en] As machine learning models penetrate critical areas like medicine, the criminal justice system, and financial markets, their opacity, which hampers humans ability to interpret most of them, has become a problem to be solved. In this work, we present a new taxonomy to classify any method, approach or strategy to deal with the problem of interpretability of machine learning models. The proposed taxonomy fills a gap in the current taxonomy frameworks regarding the subjective perception of different interpreters about the same model. To evaluate the proposed taxonomy, we have classified the contributions of some relevant scientific articles in the area.

Recent advancements in time-optimal trajectory planning, control, and state estimation for autonomous vehicles have paved the way for the emerging field of autonomous racing. In the last 5-10 years, this form of racing has become a popular and challenging testbed for autonomous driving algorithms, aiming to enhance the safety and performance of future intelligent vehicles. In autonomous racing, the main goal is to develop real-time algorithms capable of autonomously maneuvering a vehicle around a racetrack, even in the presence of moving opponents. However, as a vehicle approaches its handling limits, several challenges arise for online trajectory planning and control. The vehicle dynamics become nonlinear and hard to capture with low-complexity models, while fast re-planning and good generalization capabilities are crucial to execute optimal maneuvers in unforeseen scenarios. These challenges leave several open research questions, three of which will be addressed in this thesis. The first explores developing accurate yet computationally efficient vehicle models for online time-optimal trajectory planning. The second focuses on enhancing learning-based methods for trajectory planning, control, and state estimation, overcoming issues like poor generalization and the need for large amounts of training data. The third investigates the optimality of online-executed trajectories with simplified vehicle models, compared to offline solutions of minimum-lap-time optimal control problems using high-fidelity vehicle models. This thesis consists of four parts, each of which addresses one or more of the aforementioned research questions, in the fields of time-optimal vehicle trajectory planning, control and state estimation. The first part of the thesis presents a novel artificial race driver (ARD), which autonomously learns to drive a vehicle around an obstacle-free circuit, performing online time-optimal vehicle trajectory planning and control. The following research questions are addressed in this part: How optimal is the trajectory executed online by an artificial agent that drives a high-fidelity vehicle model, in comparison with a minimum-lap-time optimal control problem (MLT-OCP), based on the same vehicle model and solved offline? Can the artificial agent generalize to circuits and conditions not seen during training? ARD employs an original neural network with a physics-driven internal structure (PhS-NN) for steering control, and a novel kineto-dynamical vehicle model for time-optimal trajectory planning. A new learning scheme enables ARD to progressively learn the nonlinear dynamics of an unknown vehicle. When tested on a high-fidelity model of a high-performance car, ARD achieves very similar results as an MLT-OCP, based on the same vehicle model and solved offline. When tested on a 1:8 vehicle prototype, ARD achieves similar lap times as an offline optimization problem. Thanks to its physics-driven architecture, ARD generalizes well to unseen circuits and scenarios, and is robust to unmodeled changes in the vehicle’s mass. The second part of the thesis deals with online time-optimal trajectory planning for dynamic obstacle avoidance. The research questions addressed in this part are: Can time-optimal trajectory planning for dynamic obstacle avoidance be performed online and with low computational times? How optimal is the resulting trajectory? Can the planner generalize to unseen circuits and scenarios? At each planning step, the proposed approach builds a tree of time-optimal motion primitives, by performing a sampling-based exploration in a local mesh of waypoints. The novel planner is validated in challenging scenarios with multiple dynamic opponents, and is shown to be computationally efficient, to return near-time-optimal trajectories, and to generalize well to new circuits and scenarios. The third part of the thesis shows an application of time-optimal trajectory planning with optimal control and PhS-NNs in the context of autonomous parking. The research questions addressed in this part are: Can an autonomous parking framework perform fast online trajectory planning and tracking in real-life parking scenarios, such as parallel, reverse and angle parking spots, and unstructured environments? Can the framework generalize to unknown variations in the vehicle’s parameters and road adherence, and operate with measurement noise? The autonomous parking framework employs a novel penalty function for collision avoidance with optimal control, a new warm-start strategy and an original PhS-NN for steering control. The framework executes complex maneuvers in a wide range of parking scenarios, and is validated with a high-fidelity vehicle model. The framework is shown to be robust to variations in the vehicle’s mass and road adherence, and to operate with realistic measurement noise. The fourth and last part of the thesis develops novel kinematics-structured neural networks (KS-NNs) to estimate the vehicle’s lateral velocity, which is a key quantity for time-optimal trajectory planning and control. The KS-NNs are a special type of PhS-NNs: their internal structure is designed to incorporate the kinematic principles, which enhances the generalization capabilities and physical explainability. The research questions addressed in this part are: Can a neural network-based lateral velocity estimator generalize well when tested on a vehicle not used for training? Can the network’s parameters be physically explainable? The approach is validated using an open dataset with two race cars. In comparison with traditional and neural network estimators of the literature, the KS-NNs improve noise rejection, exhibit better generalization capacity, are more sample-efficient, and their structure is physically explainable.

SOTA language models have demonstrated remarkable capabilities in tackling NLP tasks they have not been explicitly trained on – given a few demonstrations of the task (few-shot learning), or even none at all (zero-shot learning). The purpose of this Master’s thesis has been to investigate practical aspects and potential applications of zero/few-shot learning in the context of text classification. This includes topics such as combined usage with active learning, automated data labeling, and interpretability. Two different methods for zero/few-shot learning have been investigated, and the results indicate that:  • Active learning can be used to marginally improve few-shot performance, but it seems to be mostly beneficial in settings with very few samples (e.g. less than 10). • Zero-shot learning can be used produce reasonable candidate labels for classes in a dataset, given knowledge of the classification task at hand.  • It is difficult to trust the predictions of zero-shot text classification without access to a validation dataset, but IML methods such as saliency maps could find usage in debugging zero-shot models. / Ledande språkmodeller har uppvisat anmärkningsvärda förmågor i att lösa NLP-problem de inte blivit explicit tränade på – givet några exempel av problemet (few-shot learning), eller till och med inga alls (zero-shot learning). Syftet med det här examensarbetet har varit att undersöka praktiska aspekter och potentiella tillämpningar av zero/few-shot learning inom kontext av textklassificering. Detta inkluderar kombinerad användning med aktiv inlärning, automatiserad datamärkning, och tolkningsbarhet. Två olika metoder för zero/few-shot learning har undersökts, och resultaten indikerar att: • Aktiv inlärning kan användas för att marginellt förbättra textklassificering med few-shot learning, men detta verkar vara mest fördelaktigt i situationer med väldigt få datapunkter (t.ex. mindre än 10). • Zero-shot learning kan användas för att hitta lämpliga etiketter för klasser i ett dataset, givet kunskap om klassifikationsuppgiften av intresse. • Det är svårt att lita på robustheten i textklassificering med zero-shot learning utan tillgång till valideringsdata, men metoder inom tolkningsbar maskininlärning såsom saliency maps skulle kunna användas för att felsöka zero-shot modeller.