To collaborate effectively, humans and AI agents need to trust each other. Communication between teammates is an essential component to achieve this, as it makes the AI system more understandable to humans. However, previous research lacks a focus on ways to build an AI agent's trust in its human teammate and, consequently, on how the AI's beliefs can be communicated to the human. As such, this study explores how real-time visual explanations of the AI agent's trust in its human teammate influence human trust and overall satisfaction. Through a user experiment (n=46) conducted on an Urban Search and Rescue simulation, integrating trust explanations was compared against a baseline containing no such information. Results show a statistically significant increase in both human trust and satisfaction when the explanations are provided, highlighting the need for further exploration into methods of communicating trust.

This thesis proposes and develops an interface and model in which advanced optimisation for general employee scheduling is made available to non-experts in computer science or optimisation. The interface teaches, guides, configures, dynamically creates a constraint programming (CP) model, iteratively improves, decreases black box properties, increases trust in the outcome, and complies with relevant European Union Artificial Intelligence law. The objective of this study is to allow a wider range of organisations to take advantage of CP techniques, with the potential to greatly improve efficiency, reduce unfairness, meet company goals, and improve employee satisfaction.

Employees are assigned to personalised shifts based on the expected demands of departments within an organisation, these are set by the domain expert in the field of employee scheduling. Next, to tailor the model to the organisation’s needs, the domain expert is guided in setting both the restrictive assumptions and priorities of shift assignment. To optimise the generated optimal schedule, it is encouraged to create and compare multiple sets of configurations and subsequent schedules. Historical demand data, organisation structure, contract information, and employee preferences are included. Multiple visual design iterations have been made, after which a working interface has been developed and improved iteratively in conjunction with experts in the field. Validations with external domain experts from various industries and organisations have shown that the interface performed effectively in supporting the objectives. Further research can be done to improve the speed of solving, implement diversity of solutions, support for highly custom constraints based on natural language, or interface reusability for other optimisation problems.

To aid the teachings of machine learning (ML), the usage of elaborative interrogative practice questions (EIPQ) is proposed to increase the long-term memory retention of said teaching. Firstly, the existing expectations of students in the current educational landscape are analyzed by taking a look at the undergraduate course present in Delft University of Technology's own Computer Science and Engineering program (CSE2510). Then, relevant theories and techniques for long-term memory retention through practice questions are introduced and applied to CSE2510 content. Finally, an experiment was carried out where roughly half of the participants made use of these newly created EIPQ, while the other half mostly used existing questions, serving as a control group (CQ). The results showed that, compared to the existing practice questions, the use of the newly created EIPQ had a profound impact on the long-term knowledge retention of the learning content. The participants who made use of EIPQ had an average retention ratio of 0.82, compared to the participants who made use of CQ, who had an average retention ratio of 0.57. Therefore, it is suggested that including EIPQ in our current educational model has favorable benefits to the students' knowledge retention of the learned content. A recommendation is made on how to carry out these methods in practice, keeping compatibility with existing learning objectives in mind.

The increasing presence of Machine Learning in all fields of study requires an improvement in how it is taught. Previous research on this topic examined how to teach ML concepts and highlighted the importance of using technology and leveraging relevant pedagogical content knowledge. It did not compare the impact of previous programming knowledge on the students' approach to solving ML problems. This paper explores the differences in implementation of 60 Machine Learning coding assignments using metrics that were determined by previous research to be a good indicator of code quality and computational thinking. By analysing the code submissions with these metrics, the results show several interesting insights about the students' use of functions, variables and the explanations of their thought process. However, results of the metrics are mostly inconclusive. The results from this study highlight the need for additional research on this topic to ensure that people with limited Computer Science knowledge are able to learn about it and implement it in their disciplines.

This study delves into machine learning (ML) education by conducting a comprehensive literature review, a targeted survey of ML lecturers in Dutch universities, and a comparative experiment. These methods aid in addressing the challenges of aligning teaching methods with the evolving nature of ML and the growing demands of the field, and fill in knowledge gaps on the success of different teaching methods in ML education. The paper investigates whether traditional methods are effective in equipping future engineers with the necessary skills for tomorrow's challenges, amidst the rapid advancement of ML and its applications. The literature review explores the range of teaching methods in ML education and not only, emphasizing a shift towards technology-enhanced and active learning approaches when teaching ML. A survey of ML lecturers explores the landscape of ML education in Dutch universities. The study investigates teaching methodologies, tools, and challenges, providing valuable insights into the evolving practices of ML instruction. Findings indicate a predominant trend towards adopting a blended approach, with lectures, projects, and group work forming core instructional methods. Virtual environments, active learning strategies, and staying informed through community engagement are highlighted. Word frequency and thematic analyses reveal key themes, emphasizing student-centric learning, practical application, and the integration of diverse teaching methods. Additionally, an experimental comparison of two teaching methods, lecture and jigsaw, sheds light on their seemingly similar efficacy when applied to the domain of ML education. The research contributes to the optimization of ML education practices, offering comprehensive insights for educators and policymakers.

This research investigates the impact of goal-oriented visualization on machine learning knowl-edge acquisition, particularly exploring its poten-tial to address procrastination in academic settings. By examining participants with no prior machine learning experience, the study employs compara-tive and correlational measures to analyze quiz per-
formance, study times, and visualization practices. While encouraging trends are observed, the small sample size emphasizes the need for further re-search to definitively establish the impact of visu-alization on academic performance. The study ad-heres to ethical guidelines, ensuring participant pri-vacy and obtaining informed consent, contributing to responsible research practices.

Intelligent agents are increasingly required to engage in collaboration with humans in the context of human-agent teams (HATs) to achieve shared goals. Interdependence is a fundamental concept in teamwork. It enables humans and robots to leverage their capabilities and collaboratively work towards a shared goal, fostering the development of trust through joint activities. Considering the great importance of trust, the effectiveness of trust repair strategies is crucial as they help mitigate the negative consequences of errors, enabling efficient collaboration between humans and robots. For this reason, the effectiveness of the trust repair strategies must be examined comprehensively by taking into consideration multiple factors, including the interdependence relationships within HATs. This paper aims to examine the impact of a mix of interdependence and independence relationships on trust violation and repair, but also on collaboration fluency. Thus, an experiment (n = 30) was conducted to study how interdependence affects trust violation, trust repair and collaboration fluency. Participants collaborated with a robot during a search and rescue mission in a simulated environment. Results show that there is a significant influence of interdependence on trust violations, but not on collaboration fluency or trust repair. Furthermore, the paper also emphasises the need for future research that investigates the effectiveness of trust repair strategies for HATs in different interdependence relationships.

ChatGPT, a cutting-edge technology based on LLM, demonstrated great potential in search tasks. While the importance and potential of ChatGPT are growing, the gap in the understanding of how users interact and engage in ChatGPT search remains open. Past research has extensively examined traditional information search, but there is a need for investigation into user behaviour and engagement in LLM contexts like ChatGPT. To address this gap, our study aims to examine the impact of ChatGPT expertise, topical expertise, and topical interest on user behaviour and user engagement in ChatGPT search as we assume these factors play an important role in shaping user interactions with ChatGPT. We conducted an experiment to investigate the answer by inviting users (N=198) via the crowdsourcing platform to communicate with the mock ChatGPT application and their interactions were recorded for subsequent analysis. Prior to and after their interaction with the mock ChatGPT application, users are requested to complete a questionnaire to gather information about their user profiles and quantify their engagement. Our finding indicates that ChatGPT expertise has a partial influence on user engagement in ChatGPT search, which may highlight the importance of AI expertise in shaping user interactions. Furthermore, our research also indicates that the Affinity for Technology Interaction (ATI) has an impact on user engagement, which underscores the importance of understanding the psychological aspects in the age of artificial intelligence. The results of this study will not only address the current knowledge gap in ChatGPT search but also provide valuable information on how to improve ChatGPT to enhance user interaction experience.

With the growing importance of teamwork in higher education, effective communication and goal congruence have become vital in improving the effectiveness of student teamwork. This study aims todesign and implement an innovative system that combines a goal-setting chatbot and an effort visualizer to facilitate effective collaboration in student teams. We found that unclear goals and lack ofengagement were the main challenges in teamwork. To address these challenges, we developed asystem that combines a chatbot and an effort visualizer designed to enhance team transparency andgoal alignment through technology. The chatbot guides users to set specific collaborative goals throughthe SMART framework. At the same time, the effort visualizer displays each member’s contribution,thus increasing accountability among team members and facilitating greater participation. The systemwe designed was validated in a creative writing task (N=84) in a collaborative scenario. The resultsshowed that either the chatbot or the effort visualizer alone positively impacted student engagementand collaboration. However, when using the chatbot and effort visualizer in combination, there neededto be more clear evidence to improve engagement and collaboration effects further. This finding suggests that we pay more attention to the characteristics of individual tools and combination strategieswhen designing collaboration enhancement tools. This study provides valuable insights into using chatbots and visualization techniques to improve team collaboration and suggests new directions for futureresearch in educational technology.

Wave Function Collapse (WFC) is a powerful generative algorithm, able to create locally-similar output based on a single example input. One of the inherent limitations of the original WFC is that it often requires users to understand its inner workings, and possibly make their own ad-hoc mods to achieve satisfactory results. Besides distracting from your creative task, this strongly reduces the algorithm's effective usefulness to a small group of technical users.

We propose a novel mixed-initiative approach to WFC, aimed at overcoming these drawbacks. These methods focus on providing intuitive control to its users, in a way that matches their usual creative workflow.

Among its main features, this approach provides (i) interactive navigation through design history, including controlled backtracking, (ii) precise manual editing of the output for direct expression of design intent, (iii) interactive manipulation of the tiles, and (iv) an initial layered approach towards post-processing algorithm output. These methods combined provide the required tools for users to tweak the global appearance of the output.

We evaluated a prototype implementation of our approach among game artists and other creative professionals, and concluded that its features were largely considered intuitive and supportive to express their creative intent.

Crowd-powered conversational systems (CPCS) solicit the wisdom of crowds to quickly respond to on-demand users' needs. The very factors that make this a viable solution ---such as the availability of diverse crowd workers on-demand--- also lead to great challenges. The ever-changing pool of online workers powering conversations with individual users makes it particularly difficult to generate contextually consistent responses from a single user's standpoint. To tackle this, prior work has employed conversational facts extracted by workers to maintain a global memory, albeit with limited success. Leveraging systematic context in affective crowdsourcing tasks has remained unexplored. Through a controlled experiment, we explored if a conversational agent, dubbed ContextBot, can provide workers with the required context on the fly for successful completion of affective support tasks in CPCS, and explore the impact of ContextBot on the response quality of workers and their interaction experience. To this end, we recruited workers (N=351) from the Prolific crowdsourcing platform and carried out a 3*3 factorial between-subjects study. Experimental conditions varied based on (i) whether or not context was elicited and informed by motivational interviewing techniques (MI, non-MI, and chat history), and (ii) different conversational entry points for workers to produce responses (early, middle, and late). Our findings showed that workers who entered the conversation earliest were more likely to produce highly consistent responses after interacting with ContextBot. Better user experience from workers was expected after they interacted with ContextBot at a late entry. We found that interacting with ContextBot through task completion did not negatively impact workers' cognitive load.

Artificial Intelligence (AI) is increasingly helping people with all kinds of tasks, due to its promising capabilities. In some tasks, an AI system by itself will take over tasks, but in other tasks, an AI system making decisions on its own would be undesired due to ethical and legal reasons. In those cases, AI can still be of help by forming human-AI teams, in which humans get advice from the AI system helping them with making their final decisions. Human-AI teams are for instance used in the medical and legal fields. One problem arises, in which instances should one trust an AI system and in which not? Trusting the AI system when it is correct and trusting yourself when you are correct, results in a high appropriate reliance. If users appropriately rely on AI systems, it is possible to achieve complementary team performance, which is better than any single teammate. However, as known from previous literature, people struggle with assessing their performance and knowing how well they perform compared to peers. When one overestimates their performance this can be because of a dual burden, due to the lack of skill they also lack the skill to accurately estimate their performance. This phenomenon is called the Dunning-Kruger Effect (DKE). This raises questions about whether the inability to estimate their own capabilities would also reflect on their assessment of the AI system its performance.

In this thesis we look at how the DKE affects (appropriate) reliance on AI systems and if so, how such effects due to the DKE can be mitigated. The effects of the DKE and possible mitigation are being tested through an empirical study (N = 249). The attempt at mitigation is done by including a tutorial intervention, which has been proved in previous research to be useful in decreasing the DKE. The tutorial intervention is aimed at revealing the weaknesses of the participant and making them aware of their miscalibrated self-estimation. Furthermore, in this thesis, the effects of revealing how the AI system makes its decisions through explainable AI (XAI) are explored. The XAI consisted of highlights from logic unit-based explanations, it should allow participants to gain more understanding of the AI advice. This thesis shows how this will affect user self-assessment and reliance on the AI system.

We found that participants who overestimate themselves tend to rely less on the AI system, compared to participants that had an accurate or underestimation of their performance. After the tutorial participants have a better calibration of their self-assessment. While the designed tutorial intervention can help participants calibrate their self-assessment, it fails to promote (appropriate) reliance. Furthermore, the logic units-based explanations did not improve accurate self-assessing or increase user (appropriate) reliance on AI systems.
This thesis shows the importance of considering cognitive biases when dealing with human-AI teams and invites more research on how to handle and mitigate the DKE in human-AI decision making.

Police officers are exposed to many potentially traumatising and stressful situations, but they do not always find the right mental health help they need. In this thesis, conversational agent Robyn is developed to help officers keep an eye on their mental well-being and find help when needed. An evaluation approach is also presented. The conversational agent applies the Stress First Aid method to determine stress indicators together with the officer.
The design of the agent followed the guidelines for Stress First Aid. The agent uses text input and output and went through an iterative design process using expert feedback. The experts provided feedback on both text and structure of the conversation. Based on their input, informal language was chosen for the presentation of the agent. The lessons for similar projects drawn from this design process are managing expectations from users, scoping broad questions, getting feedback from users on wording and not to push people when discussing suicide.
The proposed evaluation would determine if the agent is effective at improving people's awareness of their mental well-being status. This is done by comparing perceived mental well-being before and after interaction to reported mental well-being. To achieve this, a survey to measure perceived mental well-being was developed. The evaluation also provides insight in what aspects of the agent need the most improvement in future iterations.

Beat detection is an important MIR research area. Due to its growing usage in multimedia applications, the need for systematic ways to evaluate beat detectors is growing too. This research tests RhythmExtractor2013, a pipeline offered by Essentia, an open-source music analysis library used in research and industry. The annotated test samples, taken from four open-source datasets - GTZAN, Ballroom, SMC MIREX and MDB Drums, had tempo transformations (uniform, randomized, incremental and decremental tempo changes) applied to them and put to test against the aforementioned extractor. F-measure was chosen to calculate the extractor’s accuracy. The results show, that the accuracy is affected mostly by the presence of steady rhythm and drums, but also by the window size during the result calculation process, with the worst scores appearing when the samples are slowed down.

Music Information Retrieval (MIR) is a field of research that focusses on extracting information from music related data. This includes the genre of music and the beats per minute (BPM) of a song. Pipelines that extract this information from music are called feature extractors. Essentia is a library for such feature extraction. Often, the audio codec and quality is not considered in research setups within the field of MIR, while this could have an influence on the results. Therefore the main research question is "How do different audio codecs and audio quality impact genre classification and beats per minute (BPM) recognition in Essentia?". To answer this, the genre has been narrowed down to rock and the chosen audio codecs are FLAC, MP3 LAME and OGG Voribs. In collaboration with Muziekweb, a Dutch music library that collects all music that has been released in The Netherlands, it was possible to gather music files in lossless format. To degrade the audio quality, classify songs and recognize BPM, python pipelines for codec conversion, rock genre classification and BPM recognition were created an ran on this data. It has been concluded that changes in audio codec and quality have an influence on genre classification and BPM recognition in Essentia. It has not been concluded which codec and quality is best to use in the field of MIR. Further research is needed to answer this.

Working with trustworthy classifier models is important to the field of music information retrieval. However studies have shown some of the classifier models may not be as trustworthy as they appear. In this paper, we examine three of such classifiers available in the Essentia toolkit that have been evaluated using cross-validation, and measure the accuracy of these genre classifiers using cross-collection methods. We define a methodology inspired by other research in information retrieval to compare the output of the classifiers to an independent set of ground truth annotations that were the result of collaboration between the users of Last.fm. The classifiers were evaluated on 341 songs from the Muziekweb collection, and the results show that the classifiers performed worse than their cross-validation results.

The GTZAN dataset, a collection of 1000 songsspanning 10 genres, proposed by Tzanetakis hasbeen around for 20 years. In this time hundredsof researches and applications have included thisdatabase. However, there seem to be some seri-ous limitations to this dataset. There are dupli-cates, mislabellings, low audio recordings and nar-row representations of genres. This paper aimsto research the effects of both audio quality andthe content of this dataset on genre classification.A Support Vector Machine (SVM) has been usedto retrain and compare different versions of thedataset. Two experiments have been proposed inthe paper. In the first experiment, a comparison be-tween a lossless dataset of high audio quality andan mp3 version of that same dataset of a loweraudio quality have been investigated. The lowerquality dataset performed worse on the SVM clas-sifier of this size. The second experiment pro-posed a new metal dataset, based on a wider andmore balanced range of metal sub-genres. Thismetal dataset has replaced the original metal partof the GTZAN dataset. Some retrainings done thisway had a higher accuracy than the original, givingconfidence that representing a well-balanced genremight improve classification performance. Finally,it has been found that the original GTZAN classi-fier is inaccurate on audio samples outside of itsdataset, where the new retrainings done on losslessdatasets without much preprocessing seem to per-form substantially better. This last finding has notbeen verified systematically and asks for more ver-ification.

Side-channel attacks (SCA) can obtain secret information related to the private key used during encryption executed on some device by exploiting leakage in power traces produced by the device. In recent years, researchers found that a neural network (NN) can be employed to execute a powerful profiled SCA, even on targets protected with countermeasures. This paper explores the effectiveness of (1) utilising a genetic algorithm (GA) to train the weights of NNs used for SCA, (2) using NeuroEvolution of Augmenting Topologies (NEAT), which is a commonly used approach for automated NN architecture search, to construct architectures for use in SCA, and (3) \textit{Neuroevolution to Attack Side-Channel Traces Yielding Convolutional Neural Networks} (NASCTY-CNNs), a novel GA that applies genetic operators on the scale of layers rather than neurons to automatically produce CNNs for side-channel analysis. The results indicate that weight evolution and our neat-based approach are successful when attacking weak, unprotected targets. However, they do not perform well on the ASCAD data set, against which SCA methods are commonly benchmarked, suggesting these methods require significant tweaking to be effective. NASCTY can automatically produce novel NNs that achieve performance close to state-of-the-art approaches on both masked and desynchronised ASCAD traces, demonstrating that such neuroevolution methods provide a solid venue for further research.