This dissertation explores the use of Agent-based Social Simulation (ABSS) and policymodelling to support policy-making. ABSS, consisting of autonomous agents mimicking human behavior, offers a valuable alternative to traditional policy analysis methods by employing agent technology in the exploration of complex social systems and emerging behaviors. This approach allows policy-makers to perform simulated policy experiments in a safe digital environment, assessing potential adverse effects before implementation. Despite its potential, ABSS adoption in policy-making is limited. The main purpose of the dissertation is to better understand why this is the case and how the current challenges can be addressed to increase ABSS usage in policy-making.

Using a systematic review approach, six challenges in applied policy-modelling were identified: scope, politics, management, understandability, credibility, and data. It shows that upstream modelling decisions affect the rigor of model testing and highlights transparency issues like those prevalent during the COVID-19 pandemic. Furthermore, the dissertation highlights that the lack of formal accreditation and communication of model results poses a significant risk for faulty applications, which can cause societal harm. Addressing this weakness in the application-chain to increase the robustness of evidence-based policymaking is of the utmost importance.

Using a design science methodology, two artifacts were developed to address these challenges. The first includes a verification and validation protocol and an accreditation framework, facilitating independent credibility assessment of ABSS models. This design aligns model application with the scientific principle of independent review and strengthens the application-chain through quality assurance prior to application. The second artifact is a high-fidelity policy-modelling methodology employing the Institutional Grammar 2.0, ensuring systematic and transparent modelling using the case of organ donation. This methodology formalizes implemented regulations before involving subject matter experts, ensuring methodological stringency for the development of policy models. These artifacts seek to contribute to the development of realistic policy models and their responsible deployment as decision support tools within the public sector.

Falls frequently present substantial safety hazards to those who are alone, particularly the elderly. Deploying a rapid and proficient method for detecting falls is a highly effective approach to tackle this concealed peril. The majority of existing fall detection methods rely on either visual data or wearable devices, both of which have drawbacks. This research presents a multimodal approach that integrates video and audio modalities to address the issue of fall detection systems and enhances the accuracy of fall detection in challenging environmental conditions. This multimodal approach, which leverages the benefits of attention mechanism in both video and audio streams, utilizes features from both modalities through feature-level fusion to detect falls in unfavorable conditions where visual systems alone are unable to do so. We assessed the performance of our multimodal fall detection model using Le2i and UP-Fall datasets. Additionally, we compared our findings with other fall detection methods. The outstanding results of our multimodal model indicate its superior performance compared to single fall detection models.

The state space explosion, a challenge analogous to that encountered in a Petri net (PN), has constrained the extensive study of fuzzy Petri nets (FPNs). Current reasoning algorithms employing FPNs, which operate through forward, backward, and bidirectional mechanisms, are examined. These algorithms streamline the inference process by eliminating irrelevant components of the FPN. However, as the scale of the FPN grows, the complexity of these algorithms escalates sharply, posing a significant challenge for practical applications. To address the state explosion issue, this work introduces a parallel bidirectional reasoning algorithm for an FPN that utilizes reverse and decomposition strategies to optimize the implementation process. The algorithm involves hierarchically dividing a large-scale FPN into two sub-FPNs, followed by a converse operation to generate the reversal sub-FPN for the right-sub-FPN. The detailed mapping between the original and reversed FPNs is thoroughly discussed. Parallel reasoning operations are then conducted on the left-sub-FPN and the resulting reversal right-sub-FPN, with the final result derived by computing the Euclidean distance between the outcomes from the output places of the two sub-FPNs. A case study is presented to illustrate the implementation process, demonstrating the algorithm's significant enhancement of inference efficiency and substantial reduction in execution time.

The 6-minute walk test is a standardized test used in healthcare to monitor the progress of diseases affecting physical capacity and function. Inertial sensors and positioning data from wearables or smartphones allow to conduct clinical tests in patients' home environments, thereby easing the burden for patients and reducing costs for healthcare. Computation of the 6-minute walked distance requires high accuracy to be clinically useful and current consumer technology-based approaches show that noise and interference in the data often misleads algorithms used to estimate the walk distance. In this research, we are sharing a dataset of inertial and positioning information from 203 walking tests collected with users' own smartphones and the respective ground truth distances. Ground truth is measured with trundle wheels of two types, one which only provides the final distance measurement, and one which provides continuous distance measurements to also capture changes in walking speed. The tests are performed by 19 individuals, both cardiovascular patients and healthy participants. We analyse the dataset using a state-of-the-art algorithm and observe algorithm results in relation to walking features. Based on this, we elaborate on for technology development that may provide further improvements in accuracy for walk distance estimation algorithms, including how data quality and reliability can be assessed.

This paper presents a novel Hierarchical Transfer and Multi-task Learning (HTMTL) approach designed to substantially improve the performance of scene classification networks by leveraging the collective influence of diverse scene types. HTMTL is distinguished by its ability to capture the interaction between various scene types, recognizing how context information from one scene category can enhance the classification performance of another. Our method, when applied to the Places365 dataset, demonstrates a significant improvement in the network’s ability to accurately identify scene types. By exploiting these inter-scene interactions, HTMTL significantly enhances scene classification performance, making it a potent tool for advancing scene understanding and classification. Additionally, this study explores the contribution of individual tasks and task groupings on the performance of other tasks. To further validate the generality of HTMTL, we applied it to the Cityscapes dataset, where the results also show promise. This indicates the broad applicability and effectiveness of our approach across different datasets and scene types.

Walking speed and distance are usually collected when performing clinical tests such as the 6-Minute Walk Test (6MWT). Wearable devices and smartphones can help bring these tests to the home environment. However, there are difficulties in obtaining measures of distance indoors, where GPS cannot be relied on. Step counting is another even simpler form of data collection that can be obtained through digital technologies. In this work, we investigate the relationship between the step count variable and the standardised 6-Minute Walk Distance (6MWD) variable. By considering 176 6MWTs from 55 participants, we found a high correlation between ground truth distance and the number of steps taken during a test (0.83). Additionally, when considering low-quality outdoor tests, using the step count becomes significantly more reliable (MAE of 22.5m) compared to a state-of-the-art algorithm (MAE of 93.8m). We conclude that step count can be considered as a valid proxy to estimate 6MWD and a candidate approach for monitoring patients’ physical health in free-living conditions.

Deepfake and synthetic media technologies hold substantial promise for innovation in sectors such as entertainment, education, and healthcare. However, they also present significant risks, particularly in emerging economies with underdeveloped regulatory frameworks. This study proposes a comprehensive regulatory approach to balance the potential and risks of these technologies. The framework encompasses legal measures for consent and privacy, technical solutions for advanced detection tools and secure data management, and educational initiatives to enhance public awareness and digital literacy. By emphasizing collaboration with international bodies, technology companies, and community organizations, this framework aims to harness the benefits of deepfake technology while mitigating its risks, thereby fostering a safer and more ethical digital landscape in emerging economies.

This paper investigates usability aspects of a mobile application aimed at monitoring symptoms of Parkinson’s disease (PD) patients. Thirty PD patients collected data through mobile-based questionnaires and activity tasks aimed at measuring motor and non-motor symptoms for a duration of two months. We report the results about usability conducted within this study. A combination of methods consisting of the uMARS questionnaire and interviews with PD patients inform the usability aspects of the mobile application. Results indicate that the app is overall received well and is usable (median uMARS score=4). Interviews reveal usability issues related to the size of textual instructions and buttons, and to the context of use of the app, particularly when the phone is used as a sensor. These findings highlight the need of co-design and preliminary testing when developing apps for PD.

Using Artificial Intelligence-based approaches in safety-critical applications requires special attention during development. For instance, as of the beginning of 2027, European Union regulations mandate certification by a notified body for AI integration in safety-critical machinery applications. Nevertheless, AI-based approaches find application across diverse domains, enhancing system performance. Evacuation Assistance Systems used for evacuating buildings during emergencies like fires or terrorist attacks are examples in this context. In recent years, there has been a surge in research and standardization attempts to provide an assurance base for utilizing AI techniques in safety-critical applications from the technical and legislative perspectives. However, the focus is often reduced to automated driving and robotics, and many questions still need to be answered.This paper presents our research on the certification of AI-based systems. We highlight our effort in determining the relevant international standards that need to be complied with. The contribution of this paper is a certification concept for AI-based systems, where performance and reliability are crucial. The unique overview of state-of-the-art and industrial standards allows a certification attempt for this type of system. It also provides a base for future work beyond the scope of automated driving and robotics, such as assistance systems and building automation.The analytical discourse presented in this contribution justifies and highlights the mapping of standards and techniques to required functionalities and architectural components of the Evacuation Assistance System, supporting the quality and performance, system acceptance, and certification for the dedicated domain and purpose.

We evaluate deep learning architectures for rat pose estimation using a six-camera system, focusing on ResNet and EfficientNet across various depths and augmentation techniques. Among the configurations tested, ResNet 152 with default augmentation provided the best performance when employing a multi-perspective network approach in the controlled experimental setup. It reached a Root Mean Squared Error (RMSE) of 8.74, 8.78, and 9.72 pixels for the different angles. The utilization of data augmentation revealed that less altering yields better performance. We propose potential areas for future research, including further refinement of model configurations, more in-depth investigation of inference speeds, and the possibility of transferring network weights to study other species, such as mice. The findings underscore the potential for deep learning solutions to advance preclinical research in behavioral neuroscience. We suggest building on this research to introduce behavioral recogniti on based on a 3D movement reconstruction, particularly emphasizing the motoric aspects of neurodegenerative diseases. This will allow for the correlation of observable behaviors with neuronal activity, contributing to a better understanding of the brain and aiding in developing new therapeutic strategies.