Fighting videogames (also known as fighting games) are ever growing in popularity and accessibility. The isolated console experiences of 20th century gaming has been replaced by online gaming services that allow gamers to play from almost anywhere in the world with one another. This gives rise to competitive gaming on a global scale enabling them to experience fresh play styles and challenges by playing someone new. Fighting games can typically be played either as a single player experience, or against another human player, whether it is via a network or a traditional multiplayer experience. However, there are two issues with these approaches. First, the single player offering in many fighting games is regarded as being simplistic in design, making the moves by the computer predictable. Secondly, while playing against other human players can be more varied and challenging, this may not always be achievable due to the logistics involved in setting up such a bout. Game Artificial Intelligence could provide a solution to both of these issues, allowing a human player s strategy to be learned and then mimicked by the AI fighter. In this thesis, game AI techniques have been researched to provide a means of mimicking human player strategies in strategic fighting games with multiple parameters. Various techniques and their current usages are surveyed, informing the design of two separate solutions to this problem. The first solution relies solely on leveraging k nearest neighbour classification to identify which move should be executed based on the in-game parameters, resulting in decisions being made at the operational level and being fed from the bottom-up to the strategic level. The second solution utilises a number of existing Artificial Intelligence techniques, including data driven finite state machines, hierarchical clustering and k nearest neighbour classification, in an architecture that makes decisions at the strategic level and feeds them from the top-down to the operational level, resulting in the execution of moves. This design is underpinned by a novel algorithm to aid the mimicking process, which is used to identify patterns and strategies within data collated during bouts between two human players. Both solutions are evaluated quantitatively and qualitatively. A conclusion summarising the findings, as well as future work, is provided. The conclusions highlight the fact that both solutions are proficient in mimicking human strategies, but each has its own strengths depending on the type of strategy played out by the human. More structured, methodical strategies are better mimicked by the data driven finite state machine hybrid architecture, whereas the k nearest neighbour approach is better suited to tactical approaches, or even random button bashing that does not always conform to a pre-defined strategy.

Machine learning methods have been used very successfully for quite some time to recognize patterns, model correlations and generate hypotheses. However, the possibilities for weighing and evaluating the resulting models and hypotheses, and the search for alternatives and contradictions are still predominantly reserved for humans. For this purpose, the novel concept of constructivist machine learning (conML) formalizes limitations of model validity and employs constructivist learning theory to enable doubting of new and existing models with the possibility of integrating, discarding, combining, and abstracting knowledge. The present work identifies issues that impede the systems capability to abstract knowledge from generated models for tasks that lie in the domain of procedural knowledge, and proposes and implements identified solutions. To this end, the conML framework has been reimplemented in the Julia programming language and subsequently been extended. Using a synthetic dataset of impedance spectra of modeled epithelia that has previously been analyzed with an existing implementation of conML, existing and new implementations are tested for consistency and proposed algorithmic changes are evaluated with respect to changes in model generation and abstraction ability when exploring unknown data. Recommendations for specific settings and suggestions for further research are derived from the results. In terms of performance, flexibility and extensibility, the new implementation of conML in Julia provides a good starting point for further research and application of the system.:Contents Abstract . . . . . III Zusammenfassung . . . . . IV Danksagung . . . . . V Selbstständigkeitserklärung . . . . . V 1. Introduction 1.1. Research Questions . . . . . 2 2. Related Work 2.1. Hybrid AI Systems . . . . . 5 2.2. Constructivist Machine Learning (conML) . . . . . 6 2.3. Implemented Methods . . . . . 9 2.3.1. Unsupervised Machine Learning . . . . . 9 2.3.2. Supervised Machine Learning . . . . . 11 2.3.3. Supervised Feature Selection . . . . . 13 2.3.4. Unsupervised Feature Selection . . . . . 17 3. Methods and Implementation 3.1. Notable Algorithmic Changes . . . . . 19 3.1.1. Rescaling of Target Values . . . . . 19 3.1.2. ExtendedWinner Selection . . . . . 21 3.2. Package Structure . . . . . 23 3.3. Interfaces and Implementation of Specific Methods . . . . . 29 3.4. Datasets . . . . . 41 4. Results 4.1. Validation Against the conML Prototype . . . . . 43 4.2. Change in Abstraction Capability . . . . . 49 4.2.1. Influence of Target Scaling . . . . . 49 4.2.2. Influence of the Parameter kappa_p . . . . . 55 4.2.3. Influence of the Winner Selection Procedure . . . . . 61 5. Discussion 5.1. Reproduction Results . . . . . 67 5.2. Rescaling of Constructed Targets . . . . . 69 5.3. kappa_p and the Selection of Winner Models . . . . . 71 6. Conclusions 6.1. Contributions of this Work . . . . . 77 6.2. Future Work . . . . . 78 A. Julia Language Reference . . . . . 81 B. Additional Code Listings . . . . . 91 C. Available Parameters . . . . . 99 C.1. Block Processing . . . . . 105 D. Configurations Reference . . . . . 107 D.1. Unsupervised Methods . . . . . 107 D.2. Supervised Methods . . . . . 108 D.3. Feature Selection . . . . . 109 D.4. Winner Selection . . . . . 110 D.5. General Settings . . . . . 110 E. Supplemental Figures . . . . . 113 E.1. Replacing MAPE with RMSE for Z-Transform Target Scaling . . . . . 113 E.2. Combining Target Rescaling, Winner Selection and High kappa_p . . . . . 119 Bibliography . . . . . 123 List of Figures . . . . . 129 List of Listings . . . . . 133 List of Tables . . . . . 135 / Maschinelle Lernverfahren werden seit geraumer Zeit sehr erfolgreich zum Erkennen von Mustern, Abbilden von Zusammenhängen und Generieren von Hypothesen eingesetzt. Die Möglichkeiten zum Abwägen und Bewerten der entstandenen Modelle und Hypothesen, und die Suche nach Alternativen und Widersprüchen sind jedoch noch überwiegend dem Menschen vorbehalten. Das neuartige Konzept des konstruktivistischen maschinellen Lernens (conML) formalisiert dazu die Grenzen der Gültigkeit von Modellen und ermöglicht mittels konstruktivistischer Lerntheorie ein Zweifeln über neue und bestehende Modelle mit der Möglichkeit zum Integrieren, Verwerfen, Kombinieren und Abstrahieren von Wissen. Die vorliegende Arbeit identifiziert Probleme, die die Abstraktionsfähigkeit des Systems bei Aufgabenstellungen in der Prozeduralen Wissensdomäne einschränken, bietet Lösungsvorschläge und beschreibt deren Umsetzung. Das algorithmische Framework conML ist dazu in der Programmiersprache Julia reimplementiert und anschließend erweitert worden. Anhand eines synthetischen Datensatzes von Impedanzspektren modellierter Epithelien, der bereits mit einem Prototypen des conML Systems analysiert worden ist, werden bestehende und neue Implementierung auf Konsistenz geprüft und die vorgeschlagenen algorithmischen Änderungen im Hinblick auf Veränderungen beim Erzeugen von Modellen und der Abstraktionsfähigkeit bei der Exploration unbekannter Daten untersucht. Aus den Ergebnissen werden Empfehlungen zu konkreten Einstellungen sowie Vorschläge für weitere Untersuchungen abgeleitet. Die neue Implementierung von conML in Julia bietet im Hinblick auf Performanz, Flexibilität und Erweiterbarkeit einen guten Ausgangspunkt für weitere Forschung und Anwendung des Systems.:Contents Abstract . . . . . III Zusammenfassung . . . . . IV Danksagung . . . . . V Selbstständigkeitserklärung . . . . . V 1. Introduction 1.1. Research Questions . . . . . 2 2. Related Work 2.1. Hybrid AI Systems . . . . . 5 2.2. Constructivist Machine Learning (conML) . . . . . 6 2.3. Implemented Methods . . . . . 9 2.3.1. Unsupervised Machine Learning . . . . . 9 2.3.2. Supervised Machine Learning . . . . . 11 2.3.3. Supervised Feature Selection . . . . . 13 2.3.4. Unsupervised Feature Selection . . . . . 17 3. Methods and Implementation 3.1. Notable Algorithmic Changes . . . . . 19 3.1.1. Rescaling of Target Values . . . . . 19 3.1.2. ExtendedWinner Selection . . . . . 21 3.2. Package Structure . . . . . 23 3.3. Interfaces and Implementation of Specific Methods . . . . . 29 3.4. Datasets . . . . . 41 4. Results 4.1. Validation Against the conML Prototype . . . . . 43 4.2. Change in Abstraction Capability . . . . . 49 4.2.1. Influence of Target Scaling . . . . . 49 4.2.2. Influence of the Parameter kappa_p . . . . . 55 4.2.3. Influence of the Winner Selection Procedure . . . . . 61 5. Discussion 5.1. Reproduction Results . . . . . 67 5.2. Rescaling of Constructed Targets . . . . . 69 5.3. kappa_p and the Selection of Winner Models . . . . . 71 6. Conclusions 6.1. Contributions of this Work . . . . . 77 6.2. Future Work . . . . . 78 A. Julia Language Reference . . . . . 81 B. Additional Code Listings . . . . . 91 C. Available Parameters . . . . . 99 C.1. Block Processing . . . . . 105 D. Configurations Reference . . . . . 107 D.1. Unsupervised Methods . . . . . 107 D.2. Supervised Methods . . . . . 108 D.3. Feature Selection . . . . . 109 D.4. Winner Selection . . . . . 110 D.5. General Settings . . . . . 110 E. Supplemental Figures . . . . . 113 E.1. Replacing MAPE with RMSE for Z-Transform Target Scaling . . . . . 113 E.2. Combining Target Rescaling, Winner Selection and High kappa_p . . . . . 119 Bibliography . . . . . 123 List of Figures . . . . . 129 List of Listings . . . . . 133 List of Tables . . . . . 135

<p dir="ltr">In the dynamic and ever-evolving landscape of Artificial Intelligence (AI), the domain of deep learning has emerged as a pivotal force, propelling advancements across a broad spectrum of applications, notably in the intricate field of image classification. Image classification, a critical task that involves categorizing images into predefined classes, serves as the backbone for numerous cutting-edge technologies, including but not limited to, automated surveillance, facial recognition systems, and advanced diagnostics in healthcare. Despite the significant strides made in the area, the quest for models that not only excel in accuracy but also demonstrate robust generalization across varied datasets, and maintain resilience against the pitfalls of overfitting, remains a formidable challenge.</p><p dir="ltr">EfficientNetB0, a model celebrated for its optimized balance between computational efficiency and accuracy, stands at the forefront of solutions addressing these challenges. However, the nuanced complexities of datasets such as CIFAR-10, characterized by its diverse array of images spanning ten distinct categories, call for specialized adaptations to harness the full potential of such sophisticated architectures. In response, this thesis introduces an optimized version of the EffciientNetB0 architecture, meticulously enhanced with strategic architectural modifications, including the incorporation of an additional Dense layer endowed with 512 units and the strategic use of Dropout regularization. These adjustments are designed to amplify the model's capacity for learning and interpreting complex patterns inherent in the data.</p><p dir="ltr">Complimenting these architectural refinements, a nuanced two-phase training methodology is also adopted in the proposed model. This approach commences with the initial phase of training where the base model's pre-trained weights are frozen, thus leveraging the power of transfer learning to secure a solid foundational understanding. The subsequent phase of fine-tuning, characterized by the selective unfreezing of layers, meticulously calibrates the model to the intricacies of the CIFAR-10 dataset. This is further bolstered by the implementation of adaptive learning rate adjustments, ensuring the model’s training process is both efficient and responsive to the nuances of the learning curve.</p><p><br></p>

<p dir="ltr">Neurological disorders present a considerable challenge due to their variety and diagnostic complexity especially for older adults. Early prediction of the onset and ongoing assessment of the severity of these disease conditions can allow timely interventions. Currently, most of the assessment tools are time-consuming, costly, and not suitable for use in primary care. To reduce this burden, the present thesis introduces passive digital markers for different disease conditions that can effectively automate the severity assessment and risk prediction from different modalities of electronic health records (EHR). The focus of the first phase of the present study in on developing passive digital markers for the functional assessment of patients suffering from Bipolar disorder and Schizophrenia. The second phase of the study explores different architectures for passive digital markers that can predict patients at risk for dementia. The functional severity PDM uses only a single EHR modality, namely medical notes in order to assess the severity of the functioning of schizophrenia, bipolar type I, or mixed bipolar patients. In this case, the input of is a single medical note from the electronic medical record of the patient. This note is submitted to a hierarchical BERT model which classifies at-risk patients. A hierarchical attention mechanism is adopted because medical notes can exceed the maximum allowed number of tokens by most language models including BERT. The functional severity PDM follows three steps. First, a sentence-level embedding is produced for each sentence in the note using a token-level attention mechanism. Second, an embedding for the entire note is constructed using a sentence-level attention mechanism. Third, the final embedding is classified using a feed-forward neural network which estimates the impairment level of the patient. When used prior to the onset of the disease, this PDM is able to differentiate between severe and moderate functioning levels with an AUC of 76%. Disease-specific severity assessment PDMs are only applicable after the onset of the disease and have AUCs of nearly 85% for schizophrenia and bipolar patients. The dementia risk prediction PDM considers multiple EHR modalities including socio-demographic data, diagnosis codes and medical notes. Moreover, the observation period and prediction horizon are varied for a better understanding of the practical limitations of the model. This PDM is able to identify patients at risk of dementia with AUCs ranging from 70% to 92% as the observation period approaches the index date. The present study introduces methodologies for the automation of important clinical outcomes such as the assessment of the general functioning of psychiatric patients and the prediction of risk for dementia using only routine care data.</p>