In the world of autonomous driving development, simulation is used to test autonomous vehicles before real-world testing. Developing realistic simulation models has the potential to help test autonomous vehicles faster and in a larger number of realistic traffic situations and thus prepare it better before sending it out on the road. In order to stimulate the development of realistic simulation models, Waymo has introduced the Waymo open sim agents challenge where entrants are to create realistic traffic scenarios using the Waymo open motion dataset. Realism is measured through distribution matching for kinematic, interactive and map-based features, comparing a simulation to the large corpus of Waymo data. This thesis project is a take on the challenge, building on top of a transformer-based network and evaluating different refresh rates for closed-loop simulation. These are quantitatively and qualitatively analyzed and compared to two simple heuristic policies and to other submission to the challenge. The results of this project show that the transformer-based network is capable of producing some realistic scenarios, most notably at lower refresh rates. On the other side it is extremely computationally demanding, making it harder to scale than the heuristic solutions. Finally, this thesis project shows that realistic simulation is a challenging task that goes beyond achieving a high realism score as certain realistic behaviours are hard to detect using currently available metrics.

Identifying and tracking moving objects in videos is a fundamental challenge in computer vision. Although extensively studied, practical applications, especially in fields like surveying, lack focused research on tracking system performance. This study empirically compares three estimation filter-based trackers, Kalman Filter-Based Nearest-Neighbor (NN), Joint Probability Data Association Tracking (JPDA), and Gaussian Mixture Probability Hypothesis Density (GMPHD) Tracking, specifically for scenarios that include surveyor prism tracking. Simulated and real captured videos representing traffic, vegetation occlusion, and prism cluttering were analyzed to assess tracker performance in emulated real-world surveying scenarios, where precise prism tracking is crucial for efficient data collection. The results showed a similar tracking precision of 4 pixels for all three algorithms, with the highest accuracy for NN and GMPHD at 0.75 and 0.73, respectively. JPDA had many missed tracks due to failure to initiate new tracks. GMPHD experienced more track merging and ID switching than NN. Time complexity measurements revealed that NN could maintain tracking at 100 fps, while GMPHD and JPDA were slower when handling more than two and three targets, respectively. Further work is needed to assess tracking application efficiency on embedded platforms, measure power efficiency, and improve human motion modeling. 

Silicon photonics has the potential to revolutionize the electronics industry by integrating photonic functions with conventional electronic circuits on a single chip. Despite signif- icant advancements since its introduction in the 1980:s, the integration of light sources, particularly lasers, onto silicon remains a major challenge. This thesis aims for the de- velopment of a process to heterogeneously integrate III-V photonic components, such as Multi-Quantum Well (MQW) lasers and PiN photodiodes, grown on Indium Phosphide (InP) substrates, onto silicon chips using Micro Transfer Printing. The focus has primarily been on the optimization of a process flow to enable transfer and optical/electrical pumping of hybrid InP/Si Photonic Crystal Surface Emitting Lasers (PCSELs). Based on this process, continuous-wave lasing from optically pumped 1.55-μm range PCSELs fabricated on a silicon substrate was demonstrated at room temperature.

The performance of future technologies such as wireless communication, industrial automation, robotics, artificial intelligence, and intelligent sensing systems critically depends on analog-to-digital converters (ADCs), which serve as the essential interface between the analog and digital domains. State-of-the-art ADCs often employ time-interleaving to overcome the speed limitations of single-channel architectures. However, this approach introduces significant design challenges, including mismatches, crosstalk, and timing discrepancies between channels, which become more pronounced with an increasing number of channels. Therefore, enhancing the speed of individual ADC channels without compromising accuracy and power efficiency is a crucial objective. The successive-approximation-register (SAR)-assisted pipeline archi- tecture extends the application of SAR ADCs by incorporating residue amplification between two SAR-ADC stages to achieve high energy efficiency at increased sampling rates. However, the speed is limited compared to multiplying digital-to-analog converter (MDAC)-based pipeline architectures, partly due to the required high-resolution ADC per stage. To overcome this limitation, a three-stage pipeline 12-bit ADC is proposed consisting of three SAR-ADC stages with resolutions of 4, 4, and 6 bits, respectively, incorporating 1-bit interstage redundancy. The design aims to enhance speed compared to the traditional two-stage variant while maintaining high linearity and high energy efficiency. The ADC integrates two fully differential ring amplifiers (RAMPs) with 1/gm loading to achieve fast amplification times with high linearity and to provide process, voltage, and temperature robustness. Each stage employs a comparator consisting of a Floating Inverter Amplifier (FIA)-based pre-amplifier and a strong ARM latch as the second stage, utilizing dynamic biasing to improve energy efficiency while maintaining high speed. Additionally a charge redistributution digital- to-analog converter with the early reset merged capacitor switching algorithm is implemented in each stage. The work was carried out in Cadence Virtuoso at a schematic level. The total power consumption of the ADC is 3.0 mW at a sampling rate of 800 MHz. Transient simulations, in conjunction with the Fast Fourier Transform, show that when the ADC is subjected to a differential input voltage of 400 mVpp at close to Nyquist rate input frequency of 395.6 MHz, it achieves a Spurious-Free Dynamic Range (SFDR) of 90.64 dBc and a Signal-to-Noise-and-Distortion Ratio (SNDR) of 64.66 dB. While developed at a schematic level with ideal switches and Verilog-A logic, the design demonstrates significant potential for high-speed, high-linearity applications, benefiting from the reduced number of residue amplifiers due to the SAR-ADC stages.

This thesis investigates positional information within transformer based large language models which are using positional embeddings to model positional information. Within the GPT architecture, each set of pretrained weights are meant to represent something modular, the positional embeddings is one such modular piece which is meant to model the relative meaning of words based on their position in a sequence. The number of positional embeddings which the model has been trained with is what determines the model’s context length. The context length is the number of words which the model is able to infer information from at any one time. Training a GPT on a large sequence of text at a time is very expensive, thus, training a model with a very large context window can be difficult. Within this project, I modify an existing version of a pre-trained large language model, GPT-Sw3, in order to increase its context length. I take a deep dive into the learnt pattern and attributes of the positional embeddings. I also propose a method for extending the context length of a pretrained GPT with only relative light fine tuning without any major loss in ability to infer information from prior words when predicting the next token. I also look at how later architectural developments in modeling attention relate to how positional embedding represents positional information to shed light on some interesting patterns which emerge from the models’ learned patterns.

Visualising dense 3D data is a difficult task, as it is often plagued by problems of occlusion of important data. Opacity optimisation is a visualisation method which emphasises important data by modifying the opacity of geometry based its local importance, while retaining the context geometry around it where it does not occlude important parts. Approximate opacity optimisation methods operate on a per pixel basis and calculate the optimal opacity by approximating the sum of importances along a pixel, and use approximate blending methods to composite the final image. In this thesis 5 approximation methods are investigated: Fourier, Legendre, piecewise, power moments and trigonometric moments. Fourier approximation is found to have the best tradeoff between execution time and accuracy. Two rendering pipelines are also investigated, direct rendering and a method using A-buffers. The former method is found to perform better for larger datasets while the latter is quicker for smaller ones.

The method of Loci is a mnemonic where information recall is improved by visualising spatial locations. Recently, a lot of research has been done on the method of Loci in virtual reality, where participants get to memorise an ordered list of objects by placing virtual representations of the objects in virtual environments. Some studies use 2D images to represent these objects, whilst others use 3D models. There is no consensus on how objects should be represented, but this could have an impact on users memory retention. This study tests if there is a difference in memory retention from using the method of Loci with either 3D models or 2D images. A mixed reality method of Loci application was developed where objects could be represented either in 2D or 3D. Using this application, a between-subject test with 20 participants was performed to test if there was any difference in memory retention between the two modalities. The study did not find a significant difference between the 2D and 3D groups. However, the study had multiple limitations such as a limited number of participants, the limited physical space of the location where the test was performed and and the limited number of objects. A future study without the limitations in this study could find that there is a significant difference between the two, which could have a great impact on how future studies in the field are performed.

Strategy video games often present a steep learning curve for new players, as they are immediately introduced to a multitude of mechanics and concepts, potentially undermining their experience and causing early disengagement. Guided by Flow Theory-which posits that balancing challenge and skill enhances enjoyment-this study investigates Scaffolding Information Complexity (SIC), an alternative onboarding method designed to gradually introduce complexity over time. SIC was implemented in Crusader Kings III, a strategy game by Paradox Interactive, to evaluate its effectiveness in improving player experience and learning outcomes. An A/B test compared two groups: one using SIC and another without it, measuring differences in enjoyment, overall experience, and knowledge retention within similar timeframes. Contrary to expectations, results showed no statistically significant differences between the groups, with the non-SIC group reporting higher levels of knowledge acquisition. These findings raise questions about the adequacy of Flow Theory as a sole framework for understanding player enjoyment and suggest possible limitations in the method’s implementation. Further research is needed to confirm these results and determine whether they stem from theoretical or methodological factors.

Vehicle-to-Everything (V2X) communication is set to play a crucial role in the growth of In- telligent Transport System (ITS) in the coming years, driven by the increasing demand for connected and autonomous vehicles. V2X technology enables vehicles to communicate with each other and with infrastructure, enhancing road safety, reducing trac congestion, and sup- porting real-time decision-making for autonomous driving. As the adoption of V2X technologies accelerates, the need for reliable, high-performance antennas becomes crucial to ensure seamless communication in diverse and dynamic environments. This thesis presents a comprehensive analysis, design, and validation of multiple antenna typesincluding monopole, Inverted 'F' Antenna (IFA), Planar Inverted 'F' Antenna (PIFA), patch antennas specically tailored for V2X communication systems. Each antenna was designed and tested for key performance metrics such as reection coecient, gain, and radi- ation eciency, with all models operating within the 5.7 GHz to 5.925 GHz frequency range. Strategic antenna placement on vehicles was also explored, showing that optimized positioning enhances signal strength, reduces blind spots, and ensures complete 360-degree coverage. This thesis also discuss the development of SmartDisc antennas that provide seamless inte- gration with existing products, further enhancing V2X communication capabilities in roadside units and network systems. This study oers insights into optimizing antenna design and place- ment for V2X applications, contributing to the advancement of ITS. Future work will focus on real-time antenna performance measurement in vehicular environments.

Radio frequency (RF) systems play a crucial role in modern society, with applications spanning wireless communication, radar detection, and passive sensing. These systems are often required to operate across multiple frequencies to support a wider bandwidth or improve performance. Handling signals with multiple frequency components, or multi-tone signals, increases system complexity, particularly when accounting for non-ideal characteristics such as noise and intermodulation. Today, RF systems are designed using simulation software, which significantly simplifies the process of developing effective multi-tone supported systems compared to developing and testing physical prototypes. However, these complexities are reflected in the simulations, resulting in computationally intensive processes. In this thesis, a modular simulation software is developed for modelling RF systems using MATLAB. The primary goal is to create modules of key RF components – including filters, amplifiers, analogue-to-digital converters (ADCs), and digital down converters (DDC) – as well as modules of non- ideal properties such as quantisation noise, clock jitter and intermodulation products. This approach allows many different configurations of RF systems to be simulated, where each module can easily be validated separately. To evaluate the performance of the simulation software, each module was compared to its theoretically expected results. The ADC and DDC were also evaluated over a wide span of frequencies using metrics such as Signal-to- Noise Ratio (SNR), Mean Squared Error (MSE), and Effective Number of Bits (ENOB). Additionally, real-world measurements were conducted on an RF system at Saab, which were then compared to the simulation outputs. An estimation process using optimisation methods was implemented to validate parameters, such as amplitude and phase information, obtained from datasheet values of the RF components. The simulation results using the estimation process demonstrated high accuracy across all metrics. Simulation times were consistently low. Although some discrepancies remained due to limited access to component specifications and missing non-ideal properties, the results confirm the effectiveness of the simulation approach in accurately modelling RF systems.

Array antennas are widely used in wireless communication, radar detection, and radio astronomy. These types of antennas are capable of forming and steering highly directive beams in which received or transmitted signals are enhanced while suppressing signals outside the beams. Phased array antennas, in particular, can electronically steer/scan these beams by introducing phase shifts between the signals of the individual array elements, allowing for fast scanning without any mechanical movement of the antenna. The drawback of electronic steering is that the antenna’s performance decreases in terms of loss in directivity and mutual coupling as the beam is scanned away from the broadside. Phased array antennas have a limited range of scanning due to this. The range of scanning, also called the steerable sector, varies depending on the type of array that is used. The steerable sector is not universally defined for all arrays but for each individual use case. In this thesis, we define the steerable sector as the range of angles where the antenna directivity is above a specified threshold value. The main goal is to extend the steerable sector from ±45◦ to ±60◦ for a Body of Revolution array antenna. This is done by designing a dielectric radome that can be combined with the array to increase the directivity above the threshold at angles outside the steerable sector. The array operates over a broad frequency band of 6–18 GHz, which brings additional challenges to the design process. Reflections need to be removed at the material interface between the radome and surrounding air for the whole frequency band. The radome was examined through simulations at 6, 12, and 18 GHz, where it was able to fulfill the goal at 6 and 18 GHz. It was not able to completely extend the steerable sector at 12 GHz, but it was very close, as the directivity at one of the scan angles within ±60◦ was 0.14 dB below the threshold.

The evolution of mobile communications leads to the development of 6G, presenting multifaceted challenges and demanding technological shifts. It is anticipated that mm-wave and sub-THz frequency bands will be crucial to 6G. Higher operational frequencies result in shorter wavelengths, which pose challenges for the conventional design and manufacture of antenna arrays, particularly in managing grating lobes. Effective control over unwanted emissions and grating lobes is critical for the integration of future terrestrial communication systems, ensuring broad coverage and avoiding interference with non-terrestrial systems, such as satellites. Additionally, integrated radio products operating in mm-wave and beyond require highly directive antennas to counteract significant path loss while providing broad scanning and multibeam capabilities. These demands are further compounded by global economic and energy constraints, emphasizing the need for sustainable and energy-efficient radio systems with reconfigurable radiation patterns. This project addresses these challenges by developing an antenna system capable of mitigating grating lobes across a scanning range, redirecting them toward the desired direction to enhance the antenna’s directivity. The proposed solution combines a dielectric lens-based dome with a Phased Array Antenna (PAA) featuring large element spacing. This design adds quasioptical functionality in the protective radome over the PAA, which provides reconfigurability while reducing production costs. For the dielectric dome design, a simple method based on dielectric prisms and their light-redirecting ability is used to approximate its shape. As a result, the dielectric dome, created as a proof of concept for a specified array, effectively mitigated grating lobes at 0-degree and±10-degree scans, achieving reductions of up to 27.2 dB and a minimum of 10.7 dB. Additionally, with the dielectric dome, the system’s directivity improved by at least a factor of 3.1, equivalent to 4.9 dB. The lens is symmetric and static, requiring no mechanical movement or additional electromechanical components. Furthermore, the lens is constructed from dielectric materials suitable for providing both mechanical and environmental protection to the antenna system. This project sets a valuable precedent, as previous approaches to grating lobe mitigation have often resulted in a degradation of directivity in the main beam direction. The dome-array configuration explored here offers strong potential, as it effectively mitigates grating lobes and significantly enhances directivity using a simple design approach, with further optimization possible.

Multi-robot systems are essential in applications such as search and rescue, warehouse automation, and autonomous transportation, where tasks exceed the capabilities of a single agent. The effective collaboration of agents is critical, particularly in environments with communication constraints such as limited bandwidth, intermittent connectivity, and physical obstacles. This thesis aims to develop a framework for multi-robot systems that operate under these communication limitations, focusing on distributed task assignment and coordination. The proposed approach utilizes Signal Temporal Logic and Control Barrier Functions within a Quadratic Programming control framework to facilitate the indirect collaboration of agents by decomposing complex tasks into smaller sub-tasks. The framework incorporates a decentralized controller that optimizes control inputs to satisfy Signal Temporal Logic tasks while adhering to Control Barrier Function constraints. Validation is achieved through simulations and real-world experiments conducted in the Smart Mobility Lab at KTH, demonstrating that agents can achieve complex Signal Temporal Logic tasks despite limited communication. The findings contribute to enhancing the reliability and effectiveness of multi-robot systems in environments where communication is restricted.

In this work, we present a manipulation framework to reliably complete pick and place tasks using mobile bases with robotic arms aiming to enable a multi-robot system to build complex structures using different construction blocks. For picking, an eye-in-hand architecture is used and a visual servoing controller is implemented to ensure successful object tracking. The visual servoing controller is filtered by Control Barrier Functions which ensure fiducial markers in the manipulated object remain visible during approach. For placing, an extra robot with an eye-to-hand architecture is used to ensure the blocks are placed precisely, which is crucial for structural stability. Provisions are also made to allow a human in the control loop to add flexibility and fault correction capabilities to the system. A mathematical analysis of the robustness of the system with respect to errors in the pose of the camera is performed, and changes to the barrier functions are proposed to deal with them. Lastly, experiments to validate the developed framework were carried out in 6-DoF mobile manipulators.

Spiking Neural Networks (SNNs) have a biologically inspired neuron function and have demonstrated more energy efficiency than Artificial Neural Networks (ANNs) when implemented on specialized hardware. One of the challenges in using SNNs is encoding input data into the discrete spike-based representation that SNNs neurons use. This work investigates the use of SNNs with and without input rate encoding in time series classification tasks and compares their performance against conventional ANNs and non-neural network Machine Learning techniques. The study shows that the effectiveness of ANNs and SNNs architectures is strongly dependent on the dataset used, while a Decision Tree method employing Gradient Boosting showed robust performance across the experiments. While the SNNs architecture without rate encoding showed resilience with limited data availability and surpassed the performance of the ANNs in all but one of our datasets, the SNN architecture with rate encoding showed potential in specific scenarios but underperformed in achieving universal effectiveness. Cosine similarity analysis revealed a limitation of rate encoding, with small or negative numerical values encoded as spike trains containing zeros, constraining the model’s ability to recognize similarities and differences between instances belonging to the same and different classes. We also performed experiments with increased time steps in the SNNs architectures that revealed minimal enhancements in performance, suggesting the model’s effectiveness is subject to other parameters.

This thesis aims to create a basis for future research into the automation of attack graph generation. The main problem with manually creating an attack graph is that it is a costly and time-consuming endeavor. The Meta Attack Language (MAL), utilized in this thesis, is a tool that aids in the automation of this process. The related work indicates that although the process is automated to some extent, it also results in the transformation of a manual process into a software development exercise. Achieving automation is however difficult due to a widespread lack of adequate input when generating attack graphs. Thus, the main goal of this thesis has been to produce a set of requirements that need to be satisfied in order to facilitate said generation. This was done through requirements engineering, from which a prototype was designed. This prototype was then evaluated against these requirements and the ones found in the Cyber Resilience Act. The prototype in question consist of a pipeline that receives a system topology as input, from which domain-specific languages or DSLs using MAL are created. This system topology was written in Terraform, a tool used to produce infrastructure as code (IaC). The DSL in turn can be used to generate attack graphs without the involvement of a software developer. This was achieved by adding MAL toolbox (a python interface for creating attack graphs with MAL) support for generating attack graphs from machine-readable input. The evaluation of these results indicate that there is little wrong with automation tools, such as MAL, presented in this thesis. The main hurdle on the way to achieving automation and mass adoption of MAL and such tools is the lack of emphasis on software support for converting domains and topologies to machine-readable descriptions. By satisfying the various requirements proposed in this thesis however, that threshold could be substantially lowered, which is validated by the aforementioned pipeline.

Optical emission spectroscopy (OES) is a non-invasive plasma diagnostic technique that relies on the plasma’s self-emission to determine key parameters such as electron density (ne) and electron temperature (Te). Collisional radiative (CR) codes are crucial for interpreting data from OES, they provide a detailed understanding of the complex interactions within plasma. This work focuses on verification of the CoRa-He CR model, developed at the Swiss Plasma Center for helium plasmas, and its validation with measurements obtained at the Resonant Antenna Ion Device (RAID). The model has been compared with the established Goto and CraC models across a broad range of plasma conditions, showing a relative difference in state populations below 50% for states with n < 4. Additionally, experimental validation was achieved by performing OES measurements on RAID, and using ne and Te profiles obtained from Thomson scattering. The results show that CoRa-He can closely replicate observed emission line intensity ratios, especially when opacity effects are modeled with a single fitting parameter, achieving a relative difference below 60% in the central plasma region. While there is potential for further refinement, particularly in opacity modelling, the current implementation of CoRa-He demonstrates a significant step forward in developing flexible and reliable CR models for precise plasma diagnostics on the RAID device.

The Unmanned Aerial Vehicle (UAV) industry has experienced rapid growth in recent years. UAV, also known as a drone, is used for a range of applications, including aerial photography and videography, delivery of packages and medical supplies. On the other hand, drones pose a threat to physical safety, as well as the privacy of individuals and organizations, which increases the demand for solutions of drone detection and monitoring. Zadoff-chu (ZC) sequence can be the key feature for detecting drones because most commercial drones apply the ZC sequence as the synchronization sequence. However, blindly detecting the ZC sequence presents a significant challenge since the transmit frequency is unknown to the receiver. Existing studies on detecting the ZC sequence under different frequency offsets are primarily based on Long Term Evolution scenarios. However, the construction and length of the ZC sequence used in drones are different, resulting in varying outcomes. In this paper, we study the auto-correlation property of the specific ZC sequence applied by drones under different center frequency offsets, and we propose a blind drone detection and identification algorithm capable of detecting and identifying multiple drones that employ ZC sequences in their video transmission protocols. In low Signal-to-Noise (SNR) environments, even with a minimum SNR as low as -12 dB, this algorithm can still achieve a detection rate of over 99%.

This thesis investigates the transition from weekly software deployments to continuous deployments (CD) within the host company, which is an earlystage software development company. The host company was motivated to undertake this transition by the potential for productivity gains and an improved customer feedback loop, which are some of the potential benefits that CD offers in comparison to software deployments on a weekly cadence. The study addresses the technical challenges inherent in this transition and its implications for the development workflow experience, specifically for developers at the company. Notably, the company had previously attempted a shift to CD, but this endeavor was rolled back to the previous weekly deployment process due to an upsurge in errors within the product. To comprehensively explore this transition, the research employs a mixedmethods approach, a combination of qualitative and quantitative methodologies to provide a holistic view of the impact of the CD transition. Semistructured interviews were conducted with the developers at the company to gather in-depth insights into their experiences and first-hand perspectives on the transition. Additionally, a quantitative analysis of code changes over a 22- week period was performed to measure the impact on developer productivity and code stability. The post-CD transition evaluations reveal a lack of major pain points in developer productivity, with interviewees expressing no perceived increase in product errors. Despite a trend toward smaller post-transition code changes, productivity levels measured towards the end of the post-CD evaluation ultimately stabilized, aligning closely with the levels observed during the pre- CD period.

Hydropower plays an important role in the Swedish power system and is a valuable renewable energy source. Pumped storage hydroelectricity (PHES) is an application of conventional hydropower generation, where water can be pumped between two reservoirs. During low demand periods, water is pumped from a lower reservoir to an upper reservoir, where energy is stored as potential energy. This potential energy can be turned into electricity during high demand periods, by releasing the water to the lower reservoir through the turbines. The purpose of this project is to conduct a case study on investigating the profitability of implementing pumped storage hydroelectricity in the Rebnis power plant by Skellefte river. Modeling is done in a software called Spine, and economic analysis is conducted. This project includes a literature study of technical preconditions and investment costs for various solutions, and simulating different price-scenarios to assess the value of PHES. The results implicate that implementing PHES to Rebnis power plant might be profitable. The total gained revenue of with pump-operation for all simulated weeks resulted to 0,50 M€ and, a rough approximation for the whole year 2 M€. According to the NPV of 15,6 M€ and compared to the annuity of 0,91 M€ the project seems to be profitable. This project includes a variety of approximations, so to be able to assess the profitability more accurately, future work is needed. It would be interesting to study the investment costs more carefully and a longer time horizon should be used, preferably a year or longer. The environmental constraints should also be studied to be able to assess the possible environmental impacts of PHES in Rebnis.

The construction industry faces several challenges in adopting sustainable materials for building components. Engineered Wood Products (EWP) are emerging as potential alternatives to traditional materials like hollow clay blocks. This research evaluates an innovative EWP-based façade as a possible replacement for a hollow block wall in terms of acoustic and thermal performance. The study was conducted in a hotel in Guarapuava, Brazil, where acoustical measurements and thermal envelope simulations were performed. The measured Weighted Standardized Façade Level Difference (D2m,nT,w) for the existing hollow block façade was 37 dB, while the simulated data for the proposed wood façade reached 42 dB. Indoor sound insulation between rooms also improved, rising from 46 dB (measured) to 48 dB (simulated) with the EWP façade. From a thermal perspective, the thermal resistance increased from 0.50m2K/W to 1.86 m2K/W, which is more suitable for the Brazilian 1M climate zone where the building is located. This study highlights the potential of using Brazilian pine wood in façade elements.

Eco-friendly freight operations are crucial for decarbonizing the transportation sector. Systematic analysis of policy measures requires a principled modeling approach. While the commonly used model referred to as routing game considers the congestible nature of transportation facilities, exiting models fail to account for environmental factors. This paper aims at providing a mathematical framework to study strategic interaction between owners of mixed fleets comprising of both internal combustion engine vehicle (ICEV) and electric vehicle (EV) trucks. This study introduces a “green” routing game with incomplete information that models strategic interaction among multiple logistic operators. These players face a pollution tax imposed on ICEVs and a potential delayed delivery cost due to EV charging requirements with uncertainty. In contrast to existing models, this novel model captures the players' trade-off between lengthier congestion delay at charging stations as the share of EV trucks increases and higher pollution costs with increased ICEVs usage, with uncertainty determined by a latent state. We first provide equilibrium characterization and present a condition for essential uniqueness. We show that this equilibrium can be computed in a distributed manner using a gradient projection method. We then introduce a public information system that broadcasts real-time information about the latent state. Importantly, we analyze value of information for providing a condition for the public information to be beneficial. Finally, we present numerical examples to illustrate settings where environmental taxation and information dissemination can improve social welfare.

We present a moving target defense strategy to reduce the impact of stealthy sensor attacks on feedback systems. The defender periodically and randomly switches between thresholds from a discrete set to increase the uncertainty for the attacker and make stealthy attacks detectable. However, the defender does not know the exact goal of the attacker but only the prior of the possible attacker goals. Here, we model one period with a constant threshold as a Bayesian game and use the Bayesian Nash equilibrium concept to find the distribution for the choice of the threshold in that period, which takes the defender's uncertainty about the attacker into account. To obtain the equilibrium distribution, the defender minimizes its cost consisting of the cost for false alarms and the cost induced by the attack. We present a necessary and sufficient condition for the existence of a moving target defense and formulate a linear program to determine the moving target defense. Furthermore, we present a closed-form solution for the special case when the defender knows the attacker's goals. The results are numerically evaluated on a four-tank process.

Autonomous navigation problems are a real challenge when working in complex environments where uncertainty and imprecision prevail. In this thesis, we will study a robot that has to find its way through a maze. This situation can easily be modeled using a MDP (Markov Decision Process). A MDP is a reinforcement learning model in which an agent interacts stochastically with its environment. In our problem, the agent is the robot and the environment is the maze. To solve this problem, the robot must find the best actions to perform to reach its goal in the maze. HawAI.tech chose this model because of the possibility of stochastic actions. To manage the uncertainty linked to this stochasticity and solve the MDP, HawAI.tech used an algorithm called MCTS (Monte Carlo Tree Search). This algorithm achieved a success rate of almost 90%. One of the problems we faced was to give the robot access to a noisy version of its real state, rather than its actual state. In real life, sensors, like actuators, are affected by imperfections, so adding noise to the information they provide is consistent. To solve this problem, I improved our MDP representation of the problem into a POMDP (Partially Observable Markov Decision Model). In this situation, the success rate fell to 50%. Another problem was that the MCTS algorithm requires an assumption : the robot must know its exact position at all times. Once this noise had been modeled, and the assumption invalidated, the algorithm was no longer very effective, so we had to change algorithms. I transformed the MCTS algorithm into a POMCP (Partially Observable Monte Carlo Model) algorithm. To do this, I implemented a particle filter. This is a crucial element that enables the algorithm to navigate even when it has only partial and perturbed knowledge of the robot’s position. Thanks to this new feature, I was able to get back to a 93% success rate despite the noise. In addition, a new version of the problem was designed to quantify the algorithm’s handling of uncertainty through benchmarks. These benchmarks were carried out to see whether the robot takes into account the uncertainty of the information it is given for its trajectory planning. The aim was to have a robot capable of planning its trajectory while striking a balance between safety and speed.

The fan is a critical component of civil aircraft engines, converting shaft power from the core engine into thrust. Significant advancements have been made by increasing bypass ratios, thereby improving propulsion efficiency. However, larger bypass ratios also lead to increased weight and aerodynamic resistance due to larger wetted surfaces on the engine nacelle, creating a trade-off that limits fuel consumption reduction.

An alternative approach to improving efficiency is integrating the propulsor with the fuselage, allowing the fuselage boundary layer and wake to pass through the propulsor—a concept known as Boundary Layer Ingestion (BLI). This method requires less energy to accelerate the ingested flow to generate a given amount of thrust compared to freestream propulsion. To fully harness this potential, a deeper understanding of how BLI affects fan aerodynamics and transient blade loads is essential. 

To fully realize the potential of Boundary Layer Ingestion (BLI), it is essential to understand the prerequisites for designing propulsion units that perform effectively in distorted flow. In addition to efficiency, aerodynamic stability and blade vibrations must be carefully assessed.

This research consists of four interconnected components:

·       Analysis of a of a fan designed for a BLI installation

·       Design of a test object and evaluation of experimental results to verify computational tools and assess fan performance

·       Analysis of the influence on unsteady aerodynamic loads caused by distortion at the fan inlet

·       Suggestions and analysis of improved design features 

The findings indicate that propulsion efficiency can be enhanced in the studied case. Performance was evaluated for a realistic aircraft installation under relevant flight conditions, demonstrating that stability margins can be maintained. A fan with comparable performance was designed and tested at a reduced scale, with test results validating computational tools and confirming satisfactory operation across varying conditions.

A re-designed fan blade further demonstrates the feasibility of using a radial work profile to improve propulsive efficiency by compensating for the ingested boundary layer. Additionally, important new links are identified between the acoustic properties of fan blades and the unsteady blade forces generated by disturbed inlet airflow. Key design elements, including blade count and acoustic liners, are analyzed and shown to mitigate the risk of excessive blade vibrations.

Machining plays a critical role in manufacturing industries to ensure high-quality products. Therefore, understanding the machinability of different materials is crucial, as it is significantly influenced by material properties. Stainless steels are widely used owing to excellent material properties such as corrosion resistance, high strength, good ductility, etc. However, stainless steels include machining challenges due to material properties such as work hardening, lower thermal conductivity, hardness, and alloying content.

This study investigates the machinability of austenitic, martensitic, and ferritic stainless steels in terms of tool wear, tool performance, and applicability of the Colding tool life model under different cutting conditions. For this purpose, longitudinal turning tests at various cutting conditions were conducted for all stainless steel grades. Cutting tools were analysed by light optical microscope, and tool life results were used to calibrate the Colding tool life models.

The results indicate that at the same cutting conditions, tool wear and tool performance significantly vary depending on workpiece materials. Martensitic and ferritic stainless steels showed consistent tool wear behaviour, flank wear, and notch wear, respectively. Furthermore, austenitic stainless steel exhibited varying tool wear at different cutting conditions.

Additionally, martensitic and austenitic stainless steels achieved the longest tool life at 180 m/min cutting speed and 0.11 mm equivalent chip thickness, while ferritic stainless steel achieved the longest tool life at 170 m/min cutting speed and 0.193 mm equivalent chip thickness. The Colding tool life model recommended the highest cutting speed for martensitic stainless steel at equivalent chip thickness values lower than 0.3 mm but reversed this trend at equivalent chip thickness values higher than 0.3 mm. In addition, the Colding model demonstrated the highest accuracy (1.69%) for martensitic stainless steel.

After more than 10 years of abundant liquidity and easing monetary policy to limit the economic damage from the pandemic (Covid 19), central bank interest-rate start increasing to contain inflation and dry up the market from the cash excess. In this context of higher market volatility with high interest rates, the liquidity risk management is key to ensure financial stability and profitability along with lowering massive liquidity chocs among the financial institutions. This paper aims to investigate the customer churn analysis in order to know how many accounts are likely to close, and the consequences on the financial health of the bank. Data only from June 2023 are gathered, which makes it impossible to take a long-term view. The dataset is highly imbalanced: very few customers tend to churn. Therefore, resampling methods (SMOTE-NC and random undersampling) are used in order to diminish the effect of imbalance while keeping a dataset that catches different customer behaviours. Logistic Regression, Random Forest and Support Vector Machine are models used in order to predict and explain churn. We notice that, as expected, logistic regression performs well less than the two other models. Random Forest is the model that better explains churn whereas Support Vector Machine proves to be the most predictive model. These models can be improved with the incorporation of more dynamic variables (such as customer relationship) and of the economic environment.

Hur våra städer och byggda miljöer utformas påverkar livsvillkoren för alla människor, inklusive för barn. FN:s barnkonvention, som innehåller rättigheter för barn, är sedan den 1 januari 2020 inkorporerad i svensk rätt och gäller därmed som svensk lag. Inför inkorporeringen uttryckte regeringen att beaktandet av barns rättigheter och intressen var prioriterade frågor för den fysiska planeringen. Av förarbetena till inkorporeringen av barnkonventionen betonades vidare att konventionens ställning som svensk lag innebär ett förtydligande av att rättstillämpare ska beakta barns rättigheter enligt konventionen i mål och ärenden som rör barn. Inkorporeringen ansågs också utgöra en grund för ett barnrättsbaserat synsätt i all offentlig verksamhet. Några tydliga besked om hur barns rättigheter ska beaktas vid fysisk planering har dock inte lämnats från lagstiftaren eller staten. Detta väcker en rad frågor, bland annat vad barnkonventionen som svensk lag innebär för den fysiska planeringen samt hur barnkonventionen som svensk lag och plan- och bygglagen förhåller sig till varandra. 

I den här rapporten sammanfattas och redovisas de huvudsakliga resultaten från studien Barnkonventionen i kommunal planering. I studien har barnkonventionens förhållande till regelverket om planläggning enligt plan- och bygglagen studerats ur ett rättsvetenskapligt perspektiv. Studien består av tre delar. Del 1 utgör en rättslig analys av förhållandet mellan barnkonventionen som svensk lag och plan- och bygglagen med fokus på vad barnkonventionen i rättslig mening betyder för tillämpningen av plan- och bygglagen. Del 2 utgörs av en studie av planbeskrivningar till 175 detaljplaner från 15 kommuner, där fokus legat på hur kommunernas beaktat barns intressen och barns rättigheter i detaljplanerna. Del 3 är en komparativ studie där svensk och norsk lagstiftning jämförs. 

I rapporten visas att flera av rättigheterna i barnkonventionen är relevanta för den fysiska planeringen. Bland annat innehåller konventionen en rätt till vila, fritid, lek och rekreation vars förverkligande kräver att det finns lämpliga platser där barn exempelvis kan leka. Dessutom innehåller plan- och bygglagen ett antal bestämmelser som pekar ut att allmänna intressen som är viktiga för barn ska beaktas i planeringen, såsom intresset av grönområden och platser för lek. Såväl barnkonventionens bestämmelser som de allmänna intressena i plan- och bygglagen är dock allmänt hållna och lämnar ett relativt stort handlingsutrymme för kommunerna att väga olika intressen mot varandra. 

Undersökningen av regelverket visar att rättsläget är otydligt. Å ena sidan markerar inkorporeringen av barnkonventionen vikten av att barns rättigheter beaktas inom all offentlig verksamhet. Å andra sidan har barnkonventionen ännu inte i rättslig mening ansetts begränsa kommunernas handlingsutrymme. Det är med andra ord otydligt vad barnkonventionen som svensk lag faktiskt innebär för den fysiska planeringen och om konventionen innebär några skyldigheter för kommunerna. Dessa slutsatser bekräftas av studiet av planbeskrivningarna. I många beskrivningar nämns inte barn alls och i de fall barn nämns görs sällan en tydlig prövning av barns intressen och barns rättigheter. Detta tyder på att det råder osäkerhet i kommunerna om hur barns rättigheter ska beaktas. Majoriteten av de planbeskrivningar som studerats karaktäriseras vidare av ett vuxenperspektiv vilket innebär det är vuxnas perspektiv på barns rättigheter och barns intressen som ligger till grund för planeringen.

Mot denna bakgrund lämnas i rapporten ett antal medskick till såväl lagstiftaren som kommunerna. Det föreslås att lagstiftaren – om man menar allvar med att barns rättigheter är viktiga för den fysiska planeringen – bör överväga att införa bestämmelser i plan- och bygglagen som förtydligar hur barns rättigheter ska beaktas. I ett sådant reformarbete kan norsk rätt med fördel fungera som inspirationskälla. Vidare lämnas ett förslag på hur barns rättigheter kan beaktas i planeringen redan idag med utgångspunkt i principen om barnets bästa. Det konstateras också att det finns behov av ytterligare vägledning och kunskap som för samman kunskap om barns behov och intressen avseende fysisk planering och barnkonventionen som svensk lag. 

Analysis of shoulder loading can shed light on injury mechanisms, but direct measurement of loading remains challenging. Musculoskeletal simulations offer alternative estimation methods, provided they are validated. A recent open-source thoracoscapular shoulder model can reproduce scapulothoracic kinematics accurately, but its validity in estimating joint loading has been unknown. Previous attempts to use this model with  muscle redundancy solvers to estimate shoulder joint loading have been made, but whether the solutions fulfilled conditions for glenohumeral stability has been relatively little explored. The existing thoracoscapular model, moreover, does not allow any articulation in the spine, whereas muscles that span the shoulder are influenced by spinal movement.

The first aim of the thesis was to explore what degree of glenohumeral stability is adequate in a musculoskeletal model to accurately estimate shoulder joint forces. We used available kinematics and in vivo glenohumeral joint contact forces from the Orthoload dataset to evaluate the criterion validity of the proposed stability formulations. Different formulations for shoulder joint stability were introduced based on the computed direction of the joint contact force. This force was constrained to be directed into or close to the glenoid cavity, described with different geometric shapes or penalties in the muscle redundancy solver. We found that restricting the force direction towards a specified shape resulted in unrealistic force vectors that were directed along the shape borders. A less strict approach that encouraged joint contact forces to be directed centrally in the glenoid cavity estimated relatively more accurate force magnitudes and contact force directions, though some differences with the in vivo measurements still exist. 

The second aim of the thesis was to validate the use of a spine-integrated thoracoscapular shoulder (SITS) model, which includes cervical and lumbar spine articulation, to estimate shoulder biomechanics in seated activities. Specifically, aims of the second study were to evaluate the model's content validity, then to study how sitting posture can affect shoulder muscle activation and joint loading. We estimated shoulder loading during captured movements of subjects performing simple dumbbell lifting tasks in two different sitting postures—slouched and upright. We compared estimated muscle and joint loading with the rigid (locked) spine and with vertebral articulation (unlocked), and found that the customized model with an unlocked spine reproduced the actual movement more accurately. We then found that sitting postures influenced muscle activation and joint loading; compared to an upright posture, the dumbbell lateral and anterior lifting in a slouched posture involved greater glenohumeral joint movement, increased ligament lengthening, more muscle activation, and higher joint contact forces. These findings suggest that performing dumbbell lifts in a slouched posture places more load on the glenohumeral joint and increases strain on soft tissues, specifically glenohumeral ligaments.

These findings support the proposed enhanced shoulder model and stability formulations as benchmark methods for comprehensive shoulder biomechanical analysis. 

Invasive blood sampling and costly metabolic carts are the golden standard but impractical methods for tracking blood lactate levels during exercise, an important marker for performance and fatigue. This project explores a simpler, more affordable alternative which is a portable mask equipped with off-the-shelf CO₂ and total volatile organic compound (tVOC) sensors to estimate lactate-like metabolic changes in real time. By comparing exhaled breath data across different exercise intensities, the research examines whether trends in CO₂ and acetone (the primary component of tVOCs in breath) parallel rises or falls in blood lactate and workload.

Results show that while raw sensor data were influenced by factors such as breath frequency and sensor lag, an algorithmic correction to reduce the effect of varying breath frequency for CO₂ readings revealed patterns that broadly follow changes in exercise workload. Acetone readings, measured via a generic tVOC sensor, exhibited more noise and sensor contamination but still indicated potential shifts in metabolism. Although the correlation between breath-derived measures and blood lactate was moderate, partly because of limited blood lactate sampling and multi-session design, the study demonstrates the feasibility of using two low-cost sensors to gain metabolic insights. This innovation could lower barriers for sports teams, fitness enthusiasts, and clinical contexts, offering a stepping stone toward real-time, non-invasive monitoring of anaerobic metabolism and training thresholds.

This thesis investigates the comparative performance of WebGPU and WebGL in the context of 2D web-based particle systems, which lies in the area of web graphics and GPUs. WebGPU represents the next generation of web graphics API and introduces a more modern design than WebGL, including compute shaders to facilitate general-purpose computations on the GPU. However, the API is still under development and lacks full cross-platform support which makes it a relatively unexplored area of research. This study tests the performance of these APIs using particle systems entirely run on the GPU in a Google Chrome Canary browser, measuring total GPU time, render and compute time per frame, and initialization time over varying numbers of particles, sizes, and types of particles on two different GPUs, a high-end NVIDIA GeForce RTX 3080 and a lower-end GPU, Intel(R) UHD Graphics 620. The findings of the tests show that WebGPU significantly outperforms WebGL, particularly on the high-end GPU, where the updating of particle positions every frame is reduced by approximately 100 times using WebGPU over WebGL. Even on the lower-end GPU, the compute time is improved by 5 to 6 times with WebGPU. The maximum number of particles at 60 fps using WebGPU on the high-end GPU is about 37 and 20 million, depending on the type of particle. For WebGL, it is about 2.7 and 2.3 million. On the lower-end GPU, it is about 2.1 million and 398 000 for WebGPU, whereas, for WebGL, it is around 374 000 and 310 000. The results highlight the significant performance benefits that WebGPU offers and quantify the improvements compared to WebGL, which contributes to a broader understanding and insights into WebGPU as an upcoming cross-platform web graphics API.

The concept of player typology is essential in game design and marketing, as it enables the creation of more engaging and personalized gaming experiences by understanding different player types. This area has been extensively researched, highlighting the significance of player preferences and motivations in shaping game design strategies. Moreover, the interactive and immersive nature of games presents a unique opportunity to estimate player types accurately. This study explores the potential of using games as a tool to determine player types, introducing a new typology for single-player games derived from prior research. To validate this typology, a game was specifically designed to classify players into these predefined categories. Using behavioral data and questionnaire responses from 1,024 participants, the study performed correlation tests and identified several significant correlations between the participants’ in-game behavior and their gaming preferences. Participants rated the accuracy of the typology in reflecting their real-life player type, averaging 3.91 out of 5 on a Likert scale, indicating a generally positive reception but also highlighting areas for further refinement and investigation. The study also conducted a thematic analysis of participants’ feedback about the player typology and the research design, offering valuable insights for future studies. The results demonstrate that games can be a viable tool for estimating player types, accurately reflecting player preferences and motivations.

It has been shown that people behave in more aggressive ways using text-based communication. This could be because this form of communication dehumanizes us and reduces us to simple text messages. Therefore, to create more empathetic communication even when we are far apart, we should strive to make mediated experiences more like reality. With the emergence of augmented reality and virtual reality platforms such as Apple Vision Pro and Meta Quest, this is now possible in consumer hardware. This study investigates the differences in social presence between playing a game together using Apple Vision Pro with realistic anthropomorphic avatars compared to iPhone with a traditional video call. 10 participants played a game on both platforms where data would be collected using telemetry and a questionnaire. After playing the participants got interviewed. The results showed an increase in social presence on Apple Vision Pro and it was also found that the participants played slower compared to iPhone. They said that they focused more on social interactions over actually playing the game and that they used more natural interactions, like non-verbal signals. The interviews suggested that people felt more social presence because of the increased realism. Based on the results and discussion of this study, design suggestions have been created that could be applied to future products in this space.

As the global share of intermittent power in electricity production increases, driven by advances in electrification and the growth of green industries, understanding the dynamics of power trading becomes increasingly valuable. Today this is particularly true in regions like the Nordics, where wind power already contributes significantly to energy unpredictability. Power trading has emerged as a cost-efficient strategy to manage production variability, ensuring availability and stabilizing prices which impacts grid operators, consumers, producers, and policymakers alike. The global trend towards higher proportions of intermittent power suggests that insights gained in the Nordic countries could be applied to other regions in the future.

This thesis employs machine learning techniques to predict cross-border power flows using weather data from 40 weather stations in Finland and northern Sweden. The primary objective is to develop a model that predicts hourly power flows, with a secondary aim of analyzing historical modeling approaches and identifying factors influencing trade volumes and model accuracy.

The study concludes that while the proposed artificial neural networks (ANNs) and composite methods show a correlation between wind power production and cross-border flows, they currently fall short of practical application due to the complexity of the underlying problem. However, the study demonstrates that in regions heavily reliant on wind power, like the Nordics, general weather data can effectively predict wind power production, offering a cost-effective forecasting tool with relatively low data requirements.

This thesis aims to establish the ground truth of accidents and their defining characteristics through probabilistic models and feature extraction, focusing on the segmentation of multivariate time series data from e-scooter rides. The thesis evaluates the capability of Hidden Markov Model (HMM) and Multiple Regression Model with a Hidden Logistic Process (MRHLP) models against a threshold-based algorithm by analyzing rides where HMM and the threshold-based algorithm coincided. These common detected segments facilitated the use of common start indices to compute the mean and variance for both detected and undetected data points.

The findings reveal limitations in the MRHLP model's ability to precisely detect relevant segments compared to the HMM, which consistently identified significant overlaps with the threshold-based algorithm. The evaluation used the start indices of these overlapping segments to calculate the mean and variance of selected features, assessing data points 5 seconds before and after the identified segments. A linear classifier was employed to differentiate between the data points, yielding in high accuracy, recall, precision, and F1-score, with the variance of the features emerging as crucial discriminators.

Despite the limited size of the dataset, the results of this thesis indicate that the proposed method for characterizing potential accidents is promising and merits further exploration.

This project focuses on the exploration of analog Processing In Memory (PIM) architectures by leveraging the RAELLA as a baseline architecture model. This thesis extends RAELLA by incorporating real-world physical error models into the design and re-visiting the accuracy estimation. The implemented error models are integrated into the Timeloop/Accelergy simulation framework and used for accuracy evaluations of selected neural network models. We further explore the accuracy behavior under different optimization choices supported in the RAELLA architecture choice. The project addresses two primary research questions. The first question focuses on how to evaluate the accuracy of neural networks while incorporating error models, and the connection between accuracy evaluation and simulation tools, Timeloop/Accelergy. The second question is focused on the trade-offs between accuracy and hardware performance across different neural network configurations, including variations in the depth and width of neural networks. The primary research questions have been addressed with extensive ablation studies performed in a step-by-step fashion to understand the impacts of various optimization choices jointly with memristor device error models. Future studies may focus on exploring more modern memristor devices, implementing and incorporating error models for these devices in the simulator, also addressing the impacts of other non-idealities such as conductance drift on the performance of PIM architectures.

Transitioning from fossil fuels to renewable energy presents a major challenge in modern society. However, to expand the use of photovoltaic (PV) power, better prediction accuracy is required. This thesis investigated the accuracy and uncertainty of machine learning models in forecasting the hourly PV power output two days in advance, based on weather data, for a solar park situated in central Sweden. Models developed using historical observed weather data were compared against those constructed with historical forecasted weather data. The findings indicated that models trained on historical forecast data outperformed models trained on historical observed training data. This trend was consistent across all machine learning models and feature sets evaluated. Random forests and extra trees were the best performing models and including solar angles as input features improved prediction accuracy. Furthermore, relatively reliable uncertainty estimates were achievable for models trained on historical forecasted weather data using quantile gradient boosting. The study also revealed that the weather conditions of a given day significantly influenced the accuracy and uncertainty of predictions, with days featuring snow being particularly challenging to forecast accurately.

The operation of rolling stock poses distinct challenges and requirements relating to safety and reliability. Trains are intended to remain in operation across various changing factors accompanied by the large distances travelled, the variability of the weather conditions as well as the prolonged use of the trains. These factors bring with them degradation of the on-board equipment. The Train Control and Monitoring Systems record log entries, tracking the state of the train’s components. This data, paired with meteorological observations of the nearest weather stations, enable this works investigation with regards to predictive maintenance. This work builds on a foundation of unsupervised anomaly detection. It describes the combination of the data in different ways and uses deep learning models to compare the resulting anomaly score of eachlog entry. A limited set of data points with known labels is used to evaluate the models. The main metric chosen for evaluation is recall, with the aim of minimizing false negatives for safety and reliability reasons. The best models achieve 87% recall, which was an improvement compared to a baseline case. However, larger amounts of data and modelling beyond the scope of this thesis is required for application in a real world setting. Nonetheless, this thesis provides a methodology and knowledge of predictive maintenance cast as an anomaly detection problem with deliberation and comparison between two novel feature sets.

This thesis investigates the application of isolation forest outlier detection to identify potential fraudulent activities within Stockholm's public transport system, operated by Storstockholms Lokaltrafik (SL). This study employs an unsupervised machine learning model that does not rely on predefined fraud indicators, thus enabling the discovery of new and unlabelled anomalies in transaction data from travel cards and mobile app validations.

The isolation forest algorithm was chosen for its efficiency in managing large datasets and its capability to isolate anomalies based on structural differences in the data, rather than deviations from a `normal' model. The interpretability of the results was enhanced using Shapley additive explanations (SHAP) values, which showcases the influence of each feature on the decision-making process of the model.

Key findings suggest that the model effectively identified anomalous behaviors potentially indicative of fraud. However, the lack of definitive ground truth in the unsupervised setting complicated the validation of the detected anomalies. It is difficult to confirm whether some features are associated with actual payment fraud or if a user's behavior is simply unusual. Rules for identifying suspicious behavior include detecting unreasonable ticket scan counts and excessively long travel distances for certain ticket types. Scanning a ticket more than four times a day or traveling over 48 km per day with SL is highly unusual. Tickets exceeding these cutoffs warrant further investigation to identify suspicious behavior.

In recent years, there has been an increased amount of work on using deep learning for eye-tracking with the introduction of more complex neural networks that can produce accurate estimation for various eye-tracking tasks. But one issue that instead emerges is the need for large amounts of high-quality face image data to train the models. However, acquiring the data is both expensive and problematic due to privacy reasons. To address this problem, this thesis studies a way of augmenting face image data for deep-learning-based eye-tracking models using neural radiance fields (NeRFs). In particular, a more recent model called the Zip-NeRF is used to render additional views from new angles of the faces. The uncertainty of the rendered images is also studied using a Bayesian approach to ascertain the quality of them. Afterward, a series of experiments are conducted on different test sets using a residual neural network (ResNet) with 18 layers that is trained on the augmented dataset to predict 3D eyeball positions based on the face images. To demonstrate the effectiveness of the data augmentation, the results are compared with the results from a baseline that is trained with only the original images. It is found that the model trained on the augmented images is able to generalize better to face images viewed from new camera angles not seen during training. While there still are improvements to be made in the augmentation of face images, this thesis shows that there exists potential in using NeRFs to augment data in deep-learning-based eye-tracking.

This thesis investigates root finding, that is finding the first vertex, of a type of random graph, the l−dags or random recursive dags, when we add an additional Erdős-Rényi graph on the same vertex set, which represents noise. The findings show that the structure of double cycles which works for the l−dags without noise is robust to considerable noise levels. We then also explore numerically root finding algorithms for uniform attachment trees with noise, which is a current open question in the field, in order to provide some insight into future lines of research.

Generative AI is artificial intelligence that produces generative content with examples such as images, videos or audio. Most known generative AI models read the instructions given by a prompt and generate the output content based on it. However, the rapid development of generative AI carries the problem of ethical risk emergence. Examples are violent content, misinformation or the copyright of output content. Legislations are too slow to combat areas of ethical concern. This raises the question what efforts against ethical harms are generative AI companies pursuing and what implications do these strategies have. This research project is going to answer this question. Qualitative data analysis is conducted on three companies for each of three generative AI model types (text-to-image, video and audio generation types) on five areas of ethical concern (data transparency, training data, copyright, harm mitigation efforts and artist opt-out). The results show that only one generative AI company is combatting all five areas of ethical concern sufficiently and companies can do more. All companies have restricted very similar topics in their terms of service. Other ethical issues showed heterogeneous approaches between the companies. No differences or patterns in terms of approaches between companies, company values, generative AI types or other factors have been found. A notable observation is, that the less financial resources a company has, the fewer efforts against ethical complexities there are. Some companies are more transparent about their approaches and/or combat ethical issues tougher and some other companies are less transparent and/or less tough which may be attributed to various motives such as implementation costs, PR or other reasons. From the project, a conclusion is that there are no or no sufficient rules from legislations regarding these issues which means that to a high degree is up to the companies to combat ethical issues. Due to implementation costs, companies prefer fewer approaches against ethical issues in generative AI.

This project developed MeditAI, an AI-powered chatbot designed to offer guided mindfulness practices to support the well-being of older adults. Given that mindfulness practices have been shown to alleviate mental and emotional challenges, this project focuses on fine-tuning two large language models, GPT-3.5 and LLaMA-2, to generate personalized mindfulness scripts for older adults, aimed at reducing stress, fostering positive thinking, easing loneliness, and improving sleep. The models were evaluated using metrics such as BLEU, ROUGE, METEOR, and BERTScore to assess both linguistic accuracy and semantic coherence. While GPT-3.5 demonstrated superior coherence and fluency, LLaMA-2 achieved a marginally higher F1 score, highlighting its potential for tasks requiring a balance between precision and recall. These results suggest that both models have unique strengths depending on the evaluation criteria. This project offers a novel approach to supporting the mental health of older adults through personalized digital mindfulness practices. The project underscores the potential of large language models in mental health care, while also addressing the technical challenges and computational demands associated with fine-tuning. Future research could focus on optimizing model performance with larger datasets and expanding the application of digital mindfulness to other populations.

This project examines the Chan model for modeling the behavior of power transformers subjected to DC biases, focusing on electromagnetic modeling through equivalent circuits. DC biases, which can arise due to geomagnetic activity, power systems, and railway infrastructure, pose a significant threat to transformer performance by inducing DC currents that cause saturation in the transformer core. Such saturation leads to excessive heat generation, noise, and potential transformer damage. The Chan model is used to model nonlinear elements to simulate flux paths in transformers using LTSpice software to understand this behavior further. The project was carried out in collaboration with Hitachi Energy HVDC transformers. The project aims to enhance the understanding of transformer core saturation by leveraging the Chan model for nonlinear inductor cores and developing a prototype engineering tool for transformer design. The Chan model was fitted to core steel and transformer tank steel data. The EY and DY core models were implemented into the circuit modeling tool. Then the circuit models were simulated under various DC biases to observe flux behavior and saturation levels. Results from no-load and short-circuit simulations are compared to reference values that confirm the model’s validity with reasonable accuracy. The operational simulations show that flux path deviations lead to core saturation and flux taking other paths, such as through flitch plates. These findings demonstrate that the Chan model can relatively accurately simulate DC-induced transformer saturation, with short-circuit tests deviating no more than 7% and no-load tests deviating no more than 0,4%. However, more data is needed to confirm the model as a fully reliable model for wider use. Moreover, the prototype engineering tool simplifies simulation setup and analysis, enabling efficient evaluation of transformer behavior under DC bias.

We present a variational Bayesian approach to estimation of covariance matrices, which we apply on multivariate time series of financial returns. Following previous research, dynamic covariance is modeled as an inverse Wishart process constructed from independent and identically distributed centered Gaussian processes.This report details standard procedures for Gaussian process regression and prediction, alongside a review of theoretical and practical aspects of variational approximate inference. We implement a gradient-based black box variational scheme, and notably, ansatz a variational family of factorized Gaussians with kernel matrix covariance. Through experiments, we demonstrate that our proposed model is especially advantageous on shifting- and periodically correlating data, outperforming the benchmark exponential moving average in-period in terms of time-accuracy. While the framework for forecasting suffers from practical disadvantages and lacking precision on synthetic datasets, the results on financial data show modest promise.

In the world of Artificial Intelligence, Large Language Models (LLMs) are a fast growing technology. Since the launch of ChatGPT the usage of LLMs has greatly increased, both in private use and business implementations. For an LLM to work it has to be trained on vast amounts of data, but what happens when it is tasked with things that it has not been trained for? A number of problems can arise when this occurs, and the current answer for this is Retrieval Augmented Generation (RAG). RAG allows the LLM to retrieve what information it may need from varying sources and generate responses based on what it retrieved. 

This thesis aims to explore how the retrieval of information can be optimized using different retrieval methods and several retrieval enhancement methods, to explore their effects on performance using standardised metrics.

The thesis presents several methods to increase retrieval performance: Re-Ranking, Hypothetical Document Embedding (HyDE) and Query Expansion. The study also presents several metrics to evalute these methods: nDCG@10, MAP@10 and Recall@1000, with mathematical backgrounds and motivations to how they are used. The results presented in the thesis are comparable to state of the art benchmarks, however, the thesis fails to ascertain a significant increase in retrieval performance for the presented methods. It is hypothesised that this is due to shortcomings of the data used, limitations of computational hardware and time constraints which influenced what methods were chosen.

This thesis explores the generalizability of an automated deep learning algorithm with a U-Net architecture specifically tailored for human olfactory bulb (OB) segmentation. The study focuses on enhancing the performance of the segmentation tool through the implementation of targeted modifications, which are threshold adjustment and designing postprocessing constraints with the aim to successfully annotate the entire Human Connectome Project Young Adult (HCP-YA) dataset (n=1101, age 22-35 years). Extensive validation was conducted using multiple datasets to assess the algorithm's performance and its ability to generalize across different datasets with different scanning parameters. The results show a remarkably low percentage (5.8%) of missegmentations within the HCP-YA dataset and improvements in segmentation performance by reduction of over-segmentation. These enhancements do not only streamline the segmentation process but also increase the potential for using OB volume as a biomarker for neurodegenerative diseases. This thesis contributes to the field of medical image analysis by improving the performance and efficiency of OB segmentation tools, with strong implications for both research and clinical applications in diagnosing neurodegenerative disease.

The company Predictable Hiring assist companies in their recruitment process. One of the areas companies need assistance in is deciding wages to offer new recruits. Currently, there are statistical websites that offer crude means based on only a few factors. These factors do not take company specifics into consideration. In this thesis, a model is constructed based on various features that predicts an appropriate income for one type of role, but the same model can be trained for different roles, and suggests a range for appropriate minimum and maximum income offers.Various models are constructed using Linear and Polynomial regression, with and without LASSO and RIDGE regularization, Gaussian processes and Neural Networks with Ridge regularization. These models are then compared based on predictive accuracy and the best performer is selected as the model the company proceeds to use. Auxillary methods are also implemented to improve the performance of the model such as investigating outliers, evaluating magnitudes of multicolinearity, feature scaling, assessing predictive power of features and evaluating errors. The model is a functional prototype, and thus can be significantly improved upon. The company Predictable Hiring can add more data into the model upon acquisition. This thesis also contains discoveries of knowledge that can serve as a stepping stone for greater models in the future.

This thesis explores the application of generative modeling techniques to probabilistic weather forecasting, specifically through the development of DEFfusion (Direct Ensemble Forecasting with Diffusion). DEFfusion aims to address limitations in current Machine Learning Weather Prediction (MLWP) models, which tend to produce blurry forecasts and rely heavily on iterative methods. By using a diffusion model, the thesis demonstrates the potential of direct forecasting to generate more accurate and computationally efficient weather predictions. The approach involves generating an ensemble of possible future weather states from a given initial condition, utilizing a Quasi-Geostrophic (QG) model for experimental feasibility. The methodology includes training a neural network to perform dimensionality reduction via an autoencoder and employing a diffusion model to handle the probabilistic aspects of forecasting. Evaluation metrics such as Root Mean Squared Error (RMSE), Continuous Ranked Probability Score (CRPS), and Skill/Spread ratio indicate that DEFfusion offers improved performance over iterative approaches. This research contributes to advancing MLWP models by enhancing their ability to quantify uncertainty and capture extreme events, paving the way for more reliable and efficient weather prediction systems.

This thesis presents the design and optimization of a compact pulley with an integrated 360-degree swiveling carabiner, intended for high-performance training machines. The study focuses on developing an innovative mechanism that minimizes build height and friction losses while ensuring user safety and meeting the requirements of ISO 20957-1:20248. A combined approach, incorporating both manual calculations and FEM simulations, was used to evaluate key design parameters such as load distribution, friction in ball bearings, and material selection. The iterative design process enabled successive improvement of the concept proposal for the pulley and carabiner, resulting in a solution that largely meets the specification requirements. The results show that the proposed design can effectively handle both static and dynamic loads while offering intuitive use, paving the way for future improvements in the design.