In musical (jazz) improvisation, musicians that are just starting out can often feel uncomfortable when being put on the spot by their fellow players. However, when a musician is on their own when practising or leisurely playing, this prevents them from listening to fellow musicians. When a musician wants to experience some notion of co-play when they are on their own, computers and musical generative techniques may be a source of help. We study the extent to which music improvisation algorithms can facilitate such interactions by proposing an experimental framework to evaluate and compare these different algorithms. We achieve this by developing MILES ('Mixed-Initiative musicaL interactivE System'), a generic music improvisation system that allows a musician to improvise with various musical improvisation models and facilitates comparative evaluation of these models. MILES makes use of the 'trading fours' paradigm, where two or more musicians exchange four measures of solo material. We conduct experiments with novice and advanced musicians in expert-pupil and peer-to-peer settings that compare differing algorithms, as well as different variations of similar algorithms. These comparisons are based on self-assessed opinions and third-party grading and ordering of recordings, based on improvisational reciprocity and enjoyment. Symbolic music recording analysis further quantifies the interactivity between the musician and the algorithms. With this experimental setup, we are able to track familiarity and enjoyment of using music improvisation algorithms, and compare different iterations of similar music improvisation algorithms.

Social uses of virtual reality, such as collaborative virtual environments (CVEs), are showing significant increases in general adoption. In these CVEs, it is desirable for users to feel a high level of presence, which can increase collaboration effectiveness. This study investigated the effect of facial realism on presence in CVEs. Facial realism was implemented through including eye- and mouth-tracked facial expressions in an avatar representation. A controlled within-dyad experiment was performed, consisting of a dyadic interaction between participants in a CVE. We found no significant difference between a Static Face condition and either Eye Tracked or Full Tracked (eye- and mouth) facial expression conditions. Some individual items on the questionnaire were significant or marginally significant, suggesting some positive impact on the assessment of partner reactions in the Full Tracked condition, compared to Static Face. Participants also reported a lower feeling of correspondence between virtual experiences and their physical body in the Full Tracked and Eye Tracked conditions. Some evidence has been found that this experiment warrants repetition with changes discussed in this paper, which may yield significant differences with a higher sample size.

Virtual Reality's uniquely immersive nature can facilitate skill transfer between an instructor and a student. Through the use of statically transparent ghost instructors that are superimposed on the student avatar, the student can learn skills by observing the ghost instructor's movements from a first-person perspective.

This paper aims to investigate the effect that changing the transparency values of the ghost avatar based on student performance has on learning and retaining skills. A small-scale user study has been conducted to contrast these dynamically and statically transparent ghost instructors. While the group taught by statically transparent ghost instructors displayed better performance improvement between training trials, the performance drop between the training and test trials was much smaller for the group taught by dynamically transparent ghost instructors.

Lag in Virtual Reality can be devastating to the enjoyment of participants. This paper reduces the lag in a physics system by implementing a subset of the architecture proposed by Loren Peitso and Don Brutzman. This subset contains an event based physics system together with a networking system. The collision system of the physics system has been heavily reduced in size, as to speedup development of the architecture.

This architecture seems promising and handles predictable cases with almost no noticeable lag. This architecture does not handle external events well, as the lag is noticeable. The `healing' process for unexpected external events should be researched further to make this architecture viable.

Effective text input in VR is an ongoing problem with no best-agreed-upon solution. One effective solution for this is a word-gesture keyboard implemented in VR. These keyboards work by matching the user-drawn gestures to predefined gestures of all words that connect consecutive letters of the word with straight lines. We investigate matching the user input with gestures comprising of Bézier curves. We found a small improvement in writing speed and a reduction in errors while writing. The method of using complex curves for matching gestures seems promising with many possible ways for further extension.

Cognitive Impairment, commonly termed as Dementia, affects a large number of older adults. People with dementia (PwD) experience cognitive decline that impacts their ability to perform daily activities and maintain social connections. The number of PwD is expected to rise, and unfortunately, there is still no cure. A number of non drug therapies such as music and art therapy are used to stimulate neurons and slow down the progress of the disease. When symptoms become severe, full-time care may be needed, requiring a move to a nursing home for formal care. However, studies show that interaction with peers is limited in a care home, leading to loneliness and reduced quality of life. Caregivers organize group activities like singing and baking to alleviate these feelings, but their busy schedule can limit the frequency of these exercises. This begs the question: Can a robot support caregivers in the provision of group activities in a care home? This would allow caregivers more time to attend to the personal needs of the residents. It not only has the potential to reduce workload of caregivers but also creates opportunities for residents to interact with each other.

In this thesis, we propose the design of a robot which moderates a sing-along activity for a group of people living in a nursing home. An activity in the realm of music therapy is chosen since it is one of the non-drug therapies which has shown to have numerous benefits like improving memory recall, eliciting emotions and aiding relaxation. The robot session is curated to give opportunities for interaction with group members and with the robot too. While the song is playing and participants are singing along with the music, the robot performs engagement detection, and encourages those who do not seem involved.

Multiple ways are proposed to detect engagement in HRI, such as emotion recognition and estimation of head pose. These approaches have limitations in the scope of this thesis, particularly because these methods are not designed for groups of older adults residing in a care home. Given these limitations and our focus on the target group, we propose a novel engagement detection methodology, which leverages the presence of a robot and assigns an active role to it. Instead of relying on conventional passive engagement detection methods, this approach uses an interactive probing technique. The robot prompts users to perform certain actions, and whether or not they respond to it is monitored using pose estimation. This is termed robot engagement. This approach is combined with activity engagement, which indicates whether the participant is singing or not. With this hybrid technique, we aim to increase the efficacy of engagement detection. If a participant is detected as disengaged, either in the activity or in the robot, we provide personalized encouragement and motivation by addressing participants by name. Along with this, the robot also compliments those who actively participate, motivating them further...

The autonomous vehicle industry has the potential to revolutionize the future of driving, making the understanding of vehicle-driver interactions crucial as we progress towards fully autonomous systems. Advanced Driver Assistance Systems (ADAS) are integral in this evolution, bridging the gap between automated and manual driving. Research indicates that ADAS influences human driving behavior, impacting safety. While short-term behavioral effects are documented, long-term adaptations remain underexplored. Addressing this knowledge gap is vital for enhancing safety and optimizing ADAS performance. This thesis proposes a method to automatically detect long-term behavioral adaptations using a neural network and a novelty detection technique. The neural network, designed for driver identification, detects individual behavioral changes, while the novelty detection component classifies these changes as normal or abnormal. Key considerations include the need for a representative dataset, informative input features, a robust neural network architecture, and effective training. The system uses multiple Long Short-Term Memory (LSTM) layers and a novelty monitor to identify unfamiliar inputs. Evaluation using the CARLA simulator and the Intelligent Driver Model shows the approach's effectiveness in detecting behavioral adaptations, though improvements are needed for capturing subtle changes. The method's primary limitation is its lack of realism, suggesting future testing in dynamic, real-world conditions. Future work should also address defining 'normal' behavior in diverse scenarios and developing automatic parameter-setting techniques for real-time application.

In this study we introduce a different approach to the primer selection problem in the AmpliDiff [12] algorithm. Two different metrics, being the Hamming distance [2], and the Levenshtein distance [4], are used to compute sets of similar primers. This is done such that locations where mismatches between the primers and the target sequence can occur are determined by the locations of similar primers. The impact that said mismatches have on the solution set of amplicons and their respective primers is outlined. With this we show potential benefits of allowing mismatches to occur, as well as their drawbacks.

AmpliDiff is an algorithm for studying environmental samples. As the study of those DNA samples is complicated, the runtime of AmpliDiff holds back its usability. One time-consuming part is exactly solving a variant of the Set Cover problem. This paper researches whether this exact solution can be replaced by heuristics to reduce the runtime. The Iterated Local Search framework is applied as heuristic and translated onto our problem. After that, it is tested and compared to the original version of the algorithm based on the found solution size. From the results we can deduce that the heuristic version finds the same solutions as the original. It does so in a faster relative time frame as well. However, due to time constraints, the dataset that was used for testing is rather small and therefore not representative of more complex problems, which are usually solved by AmpliDiff. Thus, for future work we recommend to test the heuristics version on a larger, more representative dataset to check its viability. This paper has laid the groundwork by implementing the heuristic and showing that it effectively works for small datasets, giving reason to also try it out on more complicated problems.

Abundance estimation with the use of environmental samples has been used during the SARS-CoV-2 pandemic to identify the abundances of different lineages. AmpliDiff, an algorithm that tries to find parts of DNA that can differentiate between different input genomes was used on a SARS-CoV-2 dataset to find these amplicons. The AmpliDiff algorithm was able to run on the SARS-CoV-2 set but seemed infeasible for datasets that contain larger or more complex genomes because of the computational requirements and runtime. We introduce a new pre-processing strategy based on selecting the most differentiable coding regions and show the modifications done to AmpliDiff to make AmpliDiff work following this new method. Based on the results we conclude that the approach is promising but still requires more research to be used optimally.

This paper presents a hierarchical reinforcement learning framework for efficient robotic manipulation in sequential contact tasks. We leverage this hierarchical structure to sequentially execute behavior primitives with variable stiffness control capabilities for contact tasks. Our proposed approach relies on three key components: an action space enabling variable stiffness control, an adaptive stiffness controller for dynamic stiffness adjustments during primitive execution, and affordance coupling for efficient exploration while encouraging compliance. Through comprehensive training and evaluation, our framework learns efficient stiffness control capabilities and demonstrates improvements in learning efficiency, compositionality in primitive selection, and success rates compared to the state-of-the-art. The training environments include block lifting, door opening, object pushing, and surface cleaning. Real world evaluations further confirm the framework’s sim2real capability. This work lays the foundation for more adaptive and versatile robotic manipulation systems, with potential applications in more complex contact-based tasks.

AmpliDiff provides a method which takes a list of genomes and their lineages, and finds a set of amplicons and their primers in such a way that these amplicons can be used to differentiate between the lineages of a specific virus. While it has been shown that AmpliDiff find results comparable to whole genome sequencing for SARS-CoV-2 when looking at abundance estimations, it is not know how well it performs for other viruses, or what factors of a virus impacts the performance of the amplicons found by AmpliDiff.\\
In this paper we will be showing the effectiveness of AmpliDiff on Human monkeypox, HIV-1 and Influenza-A.
By running AmpliDiff for the three viruses mentioned above, we obtain sets of amplicons, which are used to do a lineage abundance estimation. By then comparing the estimation to the know abundance we calculate the Mean Average Error (MAE). This MAE will then be used to compare against the MAE obtained from doing a abundance estimation based on whole genome sequencing.
By comparing the amplicons against whole genome sequencing (wgs), we show that using viruses with longer genomes positively impacts the performance of the amplicons. We also show that the amount of misalignment characters added by the Multiple Sequence Alignemnt (MSA), impacts the required settings for AmpliDiff to find amplicons, and can negatively impact the MAE.
Finally, we show that AmpliDiff can be run, with some minor changes to the code base, on segmented genomes, with performance similar to that of single segment genomes.

Recommender systems are widely used to help users navigate vast content catalogs, but they often limit users to suggestions that closely match their existing preferences, creating "filter bubbles" that discourage exploration. We focus on solving this problem in the context of music recommendations, helping users discover and develop new musical tastes. We embed a knowledge graph containing expert-curated metadata, user interaction data, and audio similarity features, into a representation space where similar songs are mapped closely together. This enables the system to gradually guide users from their current preferences toward a new genre through personalized recommendations. Additionally, we apply a Bayesian active learning approach to iteratively update user preference models based on feedback, balancing exploration and exploitation to ensure user satisfaction while gathering information on the user's new preferences. We conducted a user study to evaluate the approach, demonstrating that a gradual, interactive approach outperforms directly introducing users to a new genre, increasing user engagement and their affinity toward the target genre. This research highlights the value of gradual, user-driven exploration in creating better music discovery experiences. Based on our findings, we provide recommendations for industry stakeholders and discuss opportunities for future research on targeted exploration in music recommendation.

Real-world environments require robots to continuously acquire new skills while retain-ing previously learned abilities, all without the need for clearly defined task boundaries. Storing all past data to prevent forgetting is impractical due to storage and privacy con-cerns. To address this, we propose a method that efficiently restores a robot’s proficiency in previously learned tasks over its lifespan. Using an Episodic Memory M, our approach enables experience replay during training and retrieval during testing for local fine-tuning, allowing rapid adaptation to previously encountered problems without explicit task iden-tifiers. Additionally, we introduce a selective weighting mechanism that emphasizes the most challenging segments of retrieved demonstrations, focusing local adaptation where it is most needed. This framework offers a scalable solution for lifelong learning in dy-namic, task-unaware environments, combining retrieval-based local adaptation with selec-tive weighting to enhance robot performance in open-ended scenarios.

In the task of music style transfer, the symbolic music representation based on Musical Instrument Digital Interface (MIDI) files has always been a popular research medium. By using such representation, some mature models for image style transfer can also be applied to this scenario, such as Cycle-consistent Generative Adversarial Networks (CycleGAN). However, this MIDI-based data representation is not suitable for guitar music because it does not support unique expressive information of guitar playing, such as bending, sliding, or other playing techniques. DadaGP, a dataset made up of guitar-specific format files (tablatures) and their rendered text-like tokens, enables us to perform symbolic guitar music style transfer leveraging expressive guitar playing information, and to produce playable guitar tablatures. We first adopt K-hot encoding to transform the task from sequence generation to binary classification of multiple variables, and use top-$k$ sampling to reproduce sequences from output K-hot vectors. We then propose a novel model we call CycleGMT, a CycleGAN-based model for symbolic guitar music style transfer. Finally, to mitigate the severe sparsity in the data and its resulting content loss, we adopt a skip connection between the input and output of the model, successfully achieving style-transferred music whose quality being competitive with human-composed remixes, while the musical complexity of the style-transferred music can be controlled by adjusting the value of $k$ in top-$k$ sampling.

One of the contributing factors to climate change is the release of gases, particularly carbon dioxide (CO2), which is amplified by the expanding E-commerce industry. E-commerce enterprises heavily depend on recommender systems as a means to incentivize consumers towards making product purchases. This master's thesis investigates the positive impacts of presenting information regarding carbon dioxide (CO2) emissions values on user behavior and recommendation accuracy within sustainable recommender systems. Through the creation of a new dataset, CarEmissions, this study explores whether displaying emission values influences sustainable choices in recommendations. Findings demonstrate that recommendation models trained without CO2 values consistently outperform those with CO2 values, enhancing both accuracy and greenness. This suggests that the inclusion of CO2 values introduces variability to user ratings, thereby influencing recommendation outcomes. Furthermore, this research examines correlations between user demographics, knowledge, and ratings, revealing insights into the lack of significant links. Additionally, it assesses the differences in recommendation quality between datasets with and without CO2 values, highlighting the advantages of omitting CO2 values in enhancing recommendation performance. While considering the limitations inherent to domain-specific data and convenience sampling, the thesis outlines avenues for refining data collection and exploring automated strategies for balancing recommendation accuracy and greenness. By advancing the understanding of user behavior and ethical considerations in sustainable recommender systems, this study contributes to the evolving landscape of technology-driven sustainable consumption.

As the population aged 65 and above increases, falls among these older adults emerge as a significant public health concern, leading to disabil- ities and economic burdens. Preventative strategies and personalized fall risk assessments are essential for mitigating fall risks. Human Activity Recognition in early fall risk detection by monitoring everyday ac- tivities in older adults could assess patients fall risks. However, current literature has overlooked the older adult demographic by only measuring adults younger than 65, under representing the older population. This research specifically focuses on identifying key features from head-mounted Inertial Measurement Unit (IMU) data using machine learning to classify Sit-to-Walk (STW) and Walk-to-Sit (WTS) move- ments, which are commonly associated with high risk of fall. In addition these movements can be essential in monitoring changes in performance to asses fall risk. We analyzed five activities STW, WTS, Sitting, Swing phase, and Others. Using three feature selec- tion methods (Mutual Information Gain, ANOVA, Recursive Feature Elimination) on 116 extracted fea- tures we were able to rank the features and select the top ten. The study then evaluated the accuracy of three classifiers (Logistic Regression, Random For- est, and K-Nearest Neighbor or Support Vector Ma- chine) with these features. Results indicated that the ANOVA and Random Forest classifier combination achieved the highest total accuracy of 95%, with Ran- dom Forest performing exceptionally well in STW and WTS classifications, reaching up to 81% accu- racy. Commonly selected features across all methods included the accelerometer’s maximum x-axis mea- sured and its energy in both time and frequency do- mains. This model’s performance is comparable with existing literature and validates its effectiveness in fall risk detection.

How people behave in social interactions is influenced by a multitude of factors. A large part of human communication is embedded within non-verbal communication. This type of communication is sent throughout social signals, that are embodied within low-level social cues (e.g. gaze, posture, gestures). In order for intelligent systems to seamlessly interact with humans, they need to possess some form of social intelligence. That includes expressing and recognising social signals. The field of social cue forecasting intends to predict low-level behavioral cues within social interactions, allowing systems to adapt their behavior according to the forecasted behavior of interlocutors, or synthesize human behavior on the basis of the prediction. Within social science theory, it has been established that these behavioral cues are dependent on social context, as well as individual idiosyncrasies. Under earlier work within human behavior synthesis, the latter has been mostly used, and referred to as ’style’. This work attempts to broaden the traditional view of style and proposes a model for incorporating both group, and individual-level style using a hierarchical latent variable model. To adapt to unseen groups, we incorporate this hierarchical latent structure into a meta-learning model. Introducing the hierarchical neural processes and social processes models. After testing these models on a real-world dataset containing triadic interactions, it turns out that most models fail due to posterior collapse. This prevents them from learning a useful latent representation containing semantic information with respect to forecasting future sequences of social cues. To combat this, a constant weight was assigned to a part of the loss term. However, as the issue still persists, it leaves us unable to prove whether our proposed method improves upon the baseline approach. Therefore, future work on posterior collapse in neural processes models is needed.

Social interactions in general are multifaceted and there exists an wide set of factors and events that influence them. Hence, interactions as a social phenomena have been studied by researchers in the fields of psychology and social signal processing from different stand points. The common trend in literature is to perform an in-depth study on a particular aspect of social conversations. Contrast to studying a particular aspect of social interactions, in this research, we attempt to comprehensively quantify social interactions with respect to individual experiences in the interaction, particularly focusing on spontaneous interactions which are typically non-task-directed interactions. To achieve this, we design a novel perceived measure, the perceived Conversation Quality, which intends to quantify spontaneous interactions by accounting for several aspects of spontaneous interactions, namely Rapport, Interaction Quality, Inter-personal liking and Free-for-all. Such an attempt to quantify spontaneous interactions is a substantial contribution towards building socially intelligent system to support human-human and human-robot interaction.

To quantitatively study the perceived Conversation Quality, we devised a questionnaire which measures, at the individual- and at the group- level, the perceived Conversation Quality in spontaneous interactions. Ex- isting literature in the field of social signal processing showed that behavioural features such as Turn-Taking and Bodily Coordination features are informative of several subtle social constructs like cohesion, rapport and interest-levels. Drawing inspiration from them, we extract Turn-Taking and Bodily Coordination features to model perceived Conversation Quality. Using a Logistic Regression, optimised using the Stochastic Gradi- ent Descent algorithm, we were able to predict the individual- and the group- level perceived Conversation Quality with a mean AUC of 0.76 (±0.13) and 0.96 (±0.03) respectively. From the experiments performed, we see that the Synchrony and Convergence based bodily coordination features are the best performing feature sets while predicting the individual-level perceived Conversation Quality and, the Turn-Taking based features are the best performing feature sets while predicting the group-level perceived Conversation Quality.

To further study the properties of the perceived Conversation Quality measure, we performed several statistical tests which study the effect of different social factors on the measure. From theses results, we see that the perceived Conversation Quality, both the the individual- and the group- level, decreases with the increase in number of participants in the spontaneous interaction. Moreover, the equal distribution of talk time amongst participants and the duration of their silence periods have a significant effect on the perceived Conversation Quality. Another interesting finding is that successful and unsuccessful interruptions have a positive effect and a negative effect on perceived Conversation Quality respectively. Moreover, the results also show that the time factor revealing bodily coordination features (lagged correlations and convergence) have a significant effect on both the two levels of Conversation Quality, suggesting that the time factor involved in the bodily coordination is more informative than the degree of coordination in itself. An important takeaway from the statistical test and the predictive modeling experiments is that, the two forms of perceived Conversation Quality, the individual- and the group- level, are completely different from one another with little commonality in their respective statistically significant features.