Cognitive Systems (in English)

The present dissertation goal was to research the development of an adaptive e-learning web application environment based on the user’s mental states and provide a positive user experience. The mental state that was selected was that of relaxation and concentration. These states will be calculated using a low-cost EEG device that the user will wear throughout his/her interaction with the environment. These mental states will correspond to the Available and Busy statuses of the system and will be changed automatically by the system to match with the detected mental states. Also, these status changes will also enable/disable system Notifications accordingly. Then a system evaluation will be deployed using 10 participants. The evaluation showed that such a system would indeed have a positive user experience.

In the past two decades computer chess has overcome human capabilities and efficiency in all aspects of the game. This impressive achievement has been possible, especially during the last decade, due to state-of-the-art Deep Learning methodologies that have been developed. However, since such methods perform like black-boxes, prohibiting any notion of interpretability by human users in the first place, it would be meaningful to explore the possibility of designing an explainable and cognitively efficient chess bot. In this thesis we present an efficient explainable interaction protocol accompanied by a corresponding user interface for computer chess. Moreover, we also present useful feedback from chess experts – professional players as well as chess coaches.

The present Master’s dissertation aims to research the capabilities of adapting a learning management system to the individual needs of students based on their psychological profile. The issue is viewed from the perspective of cognitive psychology, by considering personal characteristics of the users and the respective learning styles, and from the perspective of artificial intelligence, by focusing on the implementation of this concept in e-learning systems. A systematic approach to this task allows to study the relationship between the individual characteristics of students and the necessary changes in learning systems to achieve a good level of adaptability. The Master’s dissertation explores what techniques can be used to support adaptation of an e-learning system using an approach based on learning styles of its users. To achieve this goal, the following major tasks have been fulfilled in the work: 1) analysis of existing technologies for implementing user modeling and adaptation in general and in e-learning systems specifically; 2) review of personality scales that can be used as a psychological basis for defining learning styles of users; 3) analysis of methods of adaptation that can be applied at the level of planning the educational process to reflect personalization in the system based on individual characteristics. The chapters of the Master’s dissertation are dedicated respectively to three major areas: User Modeling and Adaptation, Learning Styles, and Application of AI. The User Modeling and Adaptation chapter includes a general review of user modeling, adaptation and personalization techniques with specific interest in the adaptation capabilities in e-learning platforms. The Learning Styles chapter considers psychological aspects of learning, in particular, learning styles associated with different personality types, based on a review of the literature and studies relevant to the research topic. The Application of AI chapter proposes how the suggested personality scale can be incorporated into an e-learning platform using the capabilities of Artificial Intelligence, in particular, fuzzy logic and deep learning techniques

Since the advent of the first industrial revolution, the need for machines that would help to increase production in order to fulfill market demands has increased exponentially. Industrial robots have sparked a lot of attention since then. In order to cope with industrial needs, engineers and machine designers have endeavored to construct machines that would work on the kinematics inspired by the human arm. With the developments in technology, industrial robotic arms have changed over time. Although the initial models were more hydraulic and hardwire driven, the recent robotic arms incorporate highly sophisticated mechanics, electronics, and software. With the dawn of the Fourth Industrial Revolution, industries have increased their technology benchmark and are in need of smart technology that can learn, infer, and explain their behavior. This has expanded the research in the Human Machine Interaction domain where scientists have managed to propose such systems where interacting with industrial machines has become easier. Building automation systems through no code or low code approaches has further alleviated the technology benchmarks. In this Master’s dissertation, we propose an approach under the shadow of the Human Machine Interaction domain to coach an industrial robotic arm through the PRUDENS interface that facilitates machine coaching through argumentation and machine-learning theories, which appear to be useful in monitoring the machine’s behavior and guiding it to adapt itself under exceptional settings. PRUDENS is a software tool that has been developed by the Computational Cognition Lab of the Open University of Cyprus led by Dr. Loizos Michael. We implement a real-time human-robot interaction system that facilitates machine coaching within industrial boundaries, in addition to discussing recent trends in the human-robot interaction domain and the implications of AI, ML, and argumentation techniques on it.

In this dissertation, we introduce the Cognition, Culture & Language (CCL) web ontology, based on the analytical framework of Cultural Linguistics. Cultural linguistics, which explores the relationship between cognition, culture and language, is a growing field of research. Within its paradigm, language is viewed as complex adaptive system, an emergent phenomenon arising and developing out of the varied nature of the interactions of its speakers over time. Thus, language is not simply a means of expression but a rich source of data to determine the cognitive experience of the world not only within a single human mind, but a whole community tied together by common beliefs, traditions, practices or even simply, the physical environment. If integrated with current technological advancements in Artificial Intelligence and Human-Computer Interaction, research from Cultural Linguistics could have potentially impactful applications within the domains of Education, Political discourse, AI Ethics and Social policy, just to name a few. Thus, this is a first attempt at converting this research into some suitable form of technological implementation. Ontologies for the semantic web, which are domain-specific subsets of hierarchical concepts and their relations, have widespread use in numerous AI applications due to the powerful reasoning and inference capabilities which they provide and therefore, are a natural fit for such an endeavour. CCL is a content ontology which models the core components of cultural conceptualizations. It has been assumed that the end-users of the ontology will primarily be NLP interfaces which process natural language content for use in applications within other domains. This work is inter-disciplinary, combining theories in cultural linguistics, cognition and ontology engineering for the semantic web.

The last few decades, scientists have been trying to design and produce machines capable of thinking and acting like humans “embedded” them, in a way, with cognitive abilities and human intelligence. The construction of such machines applies the research that has been conducting by several cognitive psychologists who have tried to describe the way we comprehend the information and how our cognitive functions operate and interacting together. Also, other researchers had tried to examine how humans make inferences and how these inferences are produced under a certain contexts. The results of this research will be used for the construction of a system, the Call Assistant, capable of voice interactions using itself as a personal automation system, designed for managing the phone calls. The agent should be able to learn, and to be improved, from its past interaction(s) with the user, by offering personalized solutions. Constructing the Call Assistant, we will review the hypotheses that argumentation-in the form of rules- is one of the tools which it can establish a common “language” that machines and humans can utilize when interacting through machine coaching

In an attempt to solve search-and-rescue problematics such as rescue time and difficulty in accessing certain search areas, a cognitive swarm of drones system is proposed, using artificial intelligence techniques interacting with cognitive components. The system’s various elements (drones’ cognition, pathfinding, policies, but also humans-swarm interactions) are elaborated, implemented and evaluated using a simulator custom-built for this dissertation. Evaluation outcomes show that cognitive functions can be beneficial to noncognitive drone components, and vice versa. Possible improvements are discussed.

The development of Intelligent Tutoring Systems has been a sought-after and rapidly growing field for research, especially now, with the emergence of Generative Artificial Intelligence. This study aspires to contribute to the topic of cognitive systems for education by proposing a Cognitive Architecture which encapsulates three contemporary psychological theories and constructs for learning, motivation, and self-regulation, namely the Zone of Proximal Development, Self-Determination Theory and Self-Regulated Learning by providing a theoretical blueprint for an Intelligent Tutoring Systems for children of 8 to12 years of age. The study employs surveys and experimental designs to preliminary tap into correlations between constructs of the three theories of contemporary approaches. By extracting data using instruments and cognitive tasks to the relevant population and their parents and teachers, the researcher attempts to identify any associations between items and factors and converge on a minimal set of variables and predictors, which in turn may lead to an efficient computational design model for a cognitive assistant that will employ optimal strategies for learning. However, the results suggest that more complex experimental designs may be needed to tap into the nuances of self-regulation and motivation. Finally, the study attempts to converge findings from the literature and offer a well-informed summarisation to psychologists, cognitive scientists, software architects and developers for future designs.

Deep learning models represent the cutting edge in Artificial Intelligence - AI. However, they lack one key element, explainability. Even though such models are able to perform so accurately, they act as black boxes, without letting us having any insight into how they made a prediction, and what drove them to reach it. This problem concerns the researchers for many years, as explainability is of utmost importance in order to use Deep Learning models in critical applications, such as those in medical, finance and automotive sectors. Explainable AI consists of a set of methods and processes that allow the users to understand and trust the results of the AI model, due to the fact that there is clarity in the decision making, and we can easily characterise the accuracy, the transparency and the fairness of the model, even in a complex situation. This will help AI to become more ‘responsible’ to its decisions, more trustworthy and able to help larger sectors and industries to adopt it. In recent years, researchers developed various techniques, which can identify the reasons behind the decision of a deep learning model. These techniques inspire and pave the way for new methods to be developed. For example, Class Activation Mapping produces heatmaps to highlight the region of the image which the model focused on in order to classify it. This visualisation of where the model is looking helps to identify whether the model is trustful or not. For example, a model could classify a train image by looking at the train tracks rather than the actual train. Despite the correct classification, the model might take into account wrong parts of the image, which could be a consequence of poor training. A more evolved technique which is based on Class Activation mapping, is Grad-CAM. This technique is considered class-specific, meaning that for the same image, it can produce a separate visualisation for each class which is present in it. Another interesting approach is Structured Attention Graphs (SAGs). This method is inspired from attention maps, which are popular tools for explaining the decision of deep learning models. The researchers argue that just one attention map is not enough. With SAGs, we can have a set of patches as attention maps, and we record the confidence level of the model on each one in order to evaluate how the model is impacted. This thesis will mainly focus on Grad-Class Activation mapping and Structured Attention Graphs. We will explain the procedures behind the image classification, we will benchmark the techniques and see how they apply in various datasets. We will also analyse their role in the general structure of explainable artificial intelligence. 

In process manufacturing industries and more specifically in the poultry industry which deals with live stocks there are non-deterministic factors that affect the growing performance of a flock which make the planning process of picking and adding new flocks more complicated for a human being without having supporting information. We propose a system where through human machine interaction the user will create enough mental models, arguments and reasons until to come to the final decision, to accomplish the task of planning the daily needs for processing.

The maritime industry has continuously evolved its use of Industrial Control Systems (ICS) systems to meet the growing demands for efficiency, safety, and environmental sustainability. ICS in the maritime industry highlights the evolution of technology and its integration into various aspects of maritime operations. ICS are extensively used in both ports and ships to monitor, control, and automate various operational processes. They enable seamless integration of various systems and processes, adapting them to the needs of the dynamic maritime environments. ICS and Supervisory Control and Data Acquisition (SCADA) systems in the maritime environment is symbiotic, with SCADA systems serving as a central component of the broader ICS ecosystem. SCADA systems provide centralized monitoring and control of industrial processes. In maritime applications, SCADA systems often integrate with other ICS components, such as propulsion systems, engine room automation, cargo handling equipment, and environmental monitoring systems. This integration allows for seamless communication and coordination between different systems onboard ships and within port facilities. SCADA systems offer a comprehensive suite of capabilities tailored to the unique requirements of the maritime industry. However, these systems pose significant challenges [1] [2] due to the critical nature of maritime operations and the increasing connectivity [3] of onboard and shore-based systems. They often lack modern security features and may have vulnerabilities that can be exploited by cyber attackers. Overall, cyber-threats targeting maritime SCADA can disrupt critical operations and have significant impacts on maritime operations, safety and security. Following the above, implementing an Intrusion Detection System (IDS) for Maritime SCADA systems is essential. So, the purpose of this study is to identify the threats that affect Maritime SCADA systems and develop an IDS to identify infiltration mechanisms and to simulate the results for evaluation purposes. For this purpose, first of all, a bibliographic review of internet sources (Google Scholar) was performed using the following keywords: "Ship's Cyber - defence", "SCADA security", "Maritime SCADA", "SCADA IDS", "Maritime Control Systems", "Cybersecurity ICS". It was then found that using Cyber Range environment seemed the ideal solution for this purpose. Through the Cyber Range environment, a realistic and controlled virtual scenarios that replicate the functionalities, of a maritime SCADA system both on-board and ashore was developed. Then, cyber-attacks executed that targeting these systems. The attacks included, among others, malware infections, network intrusions or denial of service attacks, with the aim of analyzing their effectiveness in a secure and controlled environment. After all the above was completed, the IDS was developed. IDS analyzed network packets, log data, and system events to identify anomalous behavior indicative of cyber threats. In this way, information obtained about the scope and severity of the detected threats. By detecting and responding to security incidents in real-time, an IDS can help organizations enhance the resilience, integrity, and security of Maritime SCADA systems in the face of evolving cyber threats. Considering the above, it is found that an IDS for Maritime SCADA systems is driven by the critical importance of safeguarding maritime infrastructure, assets, and operations from cyber threats. By detecting and responding to security incidents in real-time, an IDS helps organizations enhance the resilience, integrity, and security of Maritime SCADA systems in the face of evolving cyber threats.

The present study aimed to investigate the individual and comparative effects of Music training and Bilingualism/Multilingualism on executive functions (EFs) and working memory in healthy Greek adults. The study involved N=90 participants, who were divided into three distinct groups. The Musicians group consisted of individuals holding a degree in music and formal music training spanning over five years in various musical genres and instruments, with an average age of 33.40 years. The Bilingual/Multilingual cohort, with a mean age of 32.57 years, consisted of participants who were proficient in multiple languages, including but not limited to English, Arabic, Russian, and Italian. The control group, with an average age of 34.30 years, included people with no formal training in music or additional languages, which served as a baseline for comparison. Cognitive abilities were assessed using a series of tests: the Digit span (Wechsler, 1955), the Verbal Fluency Test (Kosmidis et al., 2004), and the Stroop Test (Stroop, 1935). These evaluations were conducted using a blended approach of in-person and digital administration to adapt to the varied environments of the participants. The research followed strict ethical protocols, ensuring informed consent, and used a detailed demographic questionnaire. Statistical analysis was performed using SPSS 25, using comparisons of means, analyses of variance, and Pearson and Spearman Rho correlations, along with hierarchical regression to dissect the cognitive performance exhibited by the different groups. Analysis of the results revealed that both Musicians and Bilinguals/Multilinguals showed higher performance on cognitive tasks compared to the control group. A comparative analysis between Musicians and Bilinguals/Multilinguals revealed differences in cognitive functions, with Musicians excelling more in certain aspects of working memory. Gender showed an effect on some cognitive tasks, while educational level showed a significant effect, especially on the results of the Stroop Test. Findings through hierarchical regression analyses reveal that music and language training advocate for the prediction of cognitive ability, with gender and educational background also playing significant roles in specific cognitive contexts. This highlights the potential of targeted music and language training to enhance cognitive abilities, suggesting avenues for future educational and cognitive development methodologies.

The aim of this Master's Dissertation is to implement an AI-based framework that currently is used today by many retail companies. This framework which incorporates conversational agents, includes an interaction between a human and a machine. Conversational Agents aim not only to provide a smoother and faster customer service for a commercial enterprise but also to increase enterprise’s business performance. At this Master’s Dissertation, the technical features that are common characteristics to most agents are examined together with the explanation of the process which agents are based in order to respond to simple questions. The implementation of the agent has been performed via the open source software Rasa.

In an age dominated by digital interconnectedness, the importance of cyber resilience cannot be overstated. The ever-evolving cybersecurity landscape demands innovative approaches to counteract the sophisticated tactics employed by malicious actors. In this context, the integration of artificial intelligence (AI) and, more specifically, generative AI tools such as chatbots, has emerged as a promising avenue for fortifying cyber defenses. This paper delves into the multifaceted relationship between generative AI chatbots and cyber resilience, seeking to answer fundamental questions regarding their contributions, integration strategies, benefits, challenges, and practical applicability.

The world becomes over time smarter, technology evolves rapidly in any scientific area and according to home electrical installations then interest is concentrating mainly in internet of things and smart systems section. Various technologies are involved in a smart facility with main purpose to reduce energy wasting costs, increase comfort and provide remote control and system information about the smart facility status from anywhere through a visualization platform that is designed for the user interaction. In an internet of things system many appliances are possible to be adapted such as lights control, motion detection, climate control, shutter and blinds control, security cameras. The amount of the connected devices to the internet is rapidly increasing, the next step is the support of the associated technologies of the web of things. (WoT). (Guinard, 2011) The main characteristic of the smart system components is the ability of the smart devices of communicating with each other in digital way. The communication between the devices applied by using smart device communication protocols. The smart devices are translating the natural world different signal to a digitized information that can be transferred through the smart device interconnected network and also through a gateway to third party system for extending the flexibility of information usage. (Gubbi, 2013, p. 1645 1660) (Pelesic, 2021) The smart devices are having a great impact according to the market, stakeholders are producing billions of devices in yearly bases for fulfil the customer needs for smart control, cost reduce and facility management. According to Statista predictions for 2025 the amount of the installed smart devices will reach the 30.9 billion devices. The increased amount of the installed smart devices will increase the facility management and control exponentially. (Vailshery, 2021) The user is involving to control more complex smart systems when increasing the smart installation. A possible solution would be the adaptation of different smart technologies that are providing the ability to make things easier to control without any technical knowledge experience. The voice assistants are having a great impact in to the market as a secondary control of a smart system. The main control of a smart system from the user is applied through a dashboard that is in most of cases compatible with mobile devices, tablets and computers. The voice assistant understands the user simple commands through its speech recognition model which is the most important part of a virtual assistant. The speech recognition system consists complex neural networks for the speech recognition creation models procedure. Convolutional Neural Networks (CNN) are better v choice because are providing more accuracy and less validation accuracy considering the comparison with Basic Neural Networks. (Patil, 2021) The user is interacting with the system with main criteria the simple control, reliability and operability. Many techniques have been applied in the past and are under development for making things easier to control for the user and the installers because these methods will increase the market benefits and the user amount because no dedicated knowledge is required for the usage of the system. The voice assistants are mainly use cognitive services and artificial intelligence for understand voice and natural language through a device or service. The voice assistants are beginning to be more popular through smart phones but also having a great impact to smart homes according to the people trend life at the beginning. People adapted and became familiar with kind of technologies for simple tasks control and information for their facility. (Uma, 2019) Automation is one of the most complex area that users are providing investments for earn time and money by reducing energy without losing their comforts. Automation are dedicated function routines that are providing features that are considering procedures that are triggered without supervision in the most of cases. Google assistant and Alexa are opening the path by making the intelligent platforms more friendly for the user. The microcontrollers, open coding sources and electronic equipment are providing support to many intelligent solutions that discovered from many scientists, engineers, student and people that is involved with the innovation. (Singh, 2019) The automation tasks are complicated, and many times are created from the KNX partners or the system administrators. The user doesn’t have the ability anytime to know which automation must apply for reaching the desired result that automation is pointing. Many automations, scripts, events and scenes can be part of the system. By taking in to account the user demands for simple control in parallel with the market increment according to statistics considering the smart devices this dissertation introduces a method for simplify internet of things control by pointing mainly in the automation area by applying cognitive, artificial intelligence techniques and natural language processing for adapting the user demands through conversational way. (Patil, 2021) The user is difficult to remember the name or the ability of any automation, script, event or scene of the smart system. The conversational ability of the user is taking in advance to create a voice assistant that can understand and point to the most significant system spot according to the user conversational description. The user conversational description will trigger a query of many complex functions that the user had to do nothing about it, without technical knowledge describes the action, the spot of interest or the point of process and the voice assistant understands and reacts by approaching one or more-point targets by taking in advance the user description classification scores and phrases similarities. The control of a complex congested ecosystem applied according to this dissertation in real time for providing the significance of the applied result by adapting cognitive services. The artificial intelligence, cognitive behavior and natural language processing was applied through google cloud services and Dialogflow Essential mechanisms for succeed the conversational patterns and system interaction with the KNX and MQTT servers. (Hager, 2021) (Raspberry, n.d.) (HomeAssistant, n.d.) The google cloud and Dialogflow essentials services became interconnected with the appliance of fulfilment webhook with the ecosystem. The google Dialogflow provides according to our knowledge until now small talk, Machine learning algorithms and is understanding the user expressions with the appliance of logical agents. (Patil, 2021) The user provides a conversational request, and the cognitive engine is classifying the request and produce an action to the system for applied in the real world through KNX and MQTT components mainly. The creation complex and congested ecosystem for experimentation purposes it is critical for receive the installation issues, the validity, the operability, the interoperability and the control things simplification by adapting a dedicated conversational voice assistant for point significant system elements. The combination of many communication protocols that are collaborating through the cloud for room control by pointing for example an HVAC system and analyze the received data is an innovation progress for the smart facilities installations. (Vanus, 2018) The experiment was placed in real time with the user to take the control of the ecosystem for providing their opinion about reducing complexity according to their technical skills and knowledge. The experimental procedure was applied by providing simple information to the user through the voice assistant and a simple advice guide for task completion. The users interacted with the system instantly and their experience is reflected to their responses according to the statistics pie charts that was collected through google questionnaire. Considering the results which arose of the conversational dedicated voice assistant proposed method, succeed to simplify things for the users. Although people according to the experiment results most of the individuals are afraid to provide trust to artificial intelligence engines for having the full control of more personal features. Considering to the system appliance improvements are needed for the future to be applied in industry with more dedicated and flexible natural language processing models and training for adapting more complex demands that are pointing to more simple descriptive commands. Considering the natural language processing modeling in parallel with the experiment, revealed that as simple the user phrase is, is more difficult for the cognitive system to classification and recognize the user demand when applied in complex headers and procedure descriptive requests. More dedicated natural language processing models is proposed for the future for adapting the user different areas of different interests according to industrial application, smart home, hospital and cities for applying cognitive system interactions.

Mental imagery is the ability to conjure sensory information in the absence of a corresponding stimulus in the immediate environment. Closely connected to the concept of visual imagination, visual mental imagery (popularly referred to as ‘the mind’s eye’) has been studied from a psychological, neuroscientific, cognitive literary perspective; the ensuing model has been successfully translated to artificial intelligence and machine learning applications. The dissertation at hand follows a multidisciplinary approach in reviewing the existing knowledge surrounding the topic of visual mental imagery, and aims to expand this research in a novel direction. Namely, the dissertation addresses the impact various positive and negative emotions may have on the ease with which mental imagery is conjured while reading fictional text, and the quality of this imagery. As a secondary purpose, this dissertation presents criticism to existing standards of measurements, to the reliance of current testing on the term ‘vividness’, and offers an alternative combination of terms to more accurately measure imagery quality (emotional intensity, detail, saturation, sharpness and latency). How emotion affects visual mental imagery is a question yet to be explored. Therefore, insight must be gained from interactions between emotion and other cognitive functions which are closely related to mental imagery. Perception has much in common with mental imagery from a phenomenological, structural and functional perspective, and thus could be considered a great predictor. Neuroimaging studies show that cortical areas associated with memory, as well as the default mode network, are active in visual mental imagery exercises. From a cognitive literary perspective, some narratological studies examine what traits in a text cause it to evoke potent imagery. All these explorations coalesce into a within-subjects experiment design whereby participants are presented with four texts, equalised for narrative factors, each evoking one of the following emotions: joy, affection (or empathy), sadness and fear. Multiple ANOVA analyses of the resulting data showed higher saturation scores in positive, but not in negative, emotions. Other variables remained below statistical significance. When data resulting from emotive vs. non-emotive reading was compared, emotion was seen to amplify saturation and sharpness, and greatly amplify intensity. However, it did not affect the level of detail or latency.

This study employed Multivariate Pattern Analysis (MVPA) to map the developmental maturation of attentional mechanisms across different age groups- adults, older children, and younger children. By using EEG data, we analysed the neural patterns that relate to attentional selection and active suppression during visual tasks that involved distractors. The findings indicated that there is a significant developmental distinction with adults having demonstrated quicker and more robust neural responses compared to the other age groups. Older children had delayed responses whilst young children showed even more delayed responses on onsets and peaks of neural activity. These findings suggest that there is indeed a developmental maturation in the neural mechanisms of attention, which aligns with traditional methods of electrophysiological research. In this study, using MVPA we leveraged spatial and temporal precision that in turn revealed specific brain regions (primarily in occipital and parietal lobes) that are activated during attentional tasks across all age groups. Adults showed concentrated brain activity, whilst children of both groups showed more widespread brain activity, which shows that there is a developmental difference in neural specialisation and efficiency. Lastly, the results showed the potential of MVPA over tradition ERP methods in providing insights in more detail regarding complex neural dynamics.

Logic and intelligence are deeply intertwined concepts, each de4ining and enhancing the other. Throughout history, humanity has continually seek to expand the limits of its cognitive and mental capabilities. Logic, both as a concept and a science, has frequently served as a springboard and a tool for intellectual exploration and mental development. In parallel, arti4icial intelligence has evolved by integrating two primary approaches: symbolic systems (knowledge-based systems) and neural networks (connectionist systems). This integration has given rise to the novel 4ield of Neurosymbolic AI, which merges these approaches in order to create more robust and versatile AI systems. The aim of the current study, is to explore the Neurosymbolic AI approach and it's relevant potential applications. Speci4ically, the current study is deployed into three distinct but conceptually coherent parts, in which we examine the theoretical foundations, methodological advancements, and practical applications of Neurosymbolic AI novel approach, as follows: In the 4irst part of my thesis, I present the foundational pillars of Neurosymbolic AI, focusing on Mathematical Logic and Arti4icial Neural Networks. Speci4ically, I delve into the theory of Propositional and First-Order Logic and explaining some of their most important properties. Following this, I provide a comprehensive overview of Arti4icial Neural Networks (ANNs), speci4ically, regarding Feed Forward Networks, highlighting their structure, function, and signi4icance in modern AI. To underline the importance of enhancing and advancing AI methods, I conclude this section with examples of adversarial attacks that scienti4ically compromise the robustness and accuracy of ANNs and Deep Neural Networks (DNNs). The second part of my thesis focus directly into the core subject of Neurosymbolic AI, where I explore and formalize this emerging very promising 4ield as a cohesive mathematical formal theory. Neurosymbolic AI aims to integrate the strengths of symbolic reasoning, derived from logic, with the learning capabilities of neural networks. I discuss the theoretical framework that underpins this integration, examining how it can potentially address the limitations inherent in purely symbolic or neural approaches. This section is pivotal for setting the foundations for the experimental part of the study. In the 4inal part of my thesis, I present the experimental study which serves as a practical demonstration of the theories discussed before. This study involves implementing a hybrid model that combines symbolic logic with neural networks, by embedding a fundamental mathematical axiom, that of the identity element into an augmented hybrid-knowledgedatabase, where we aim to train our classi4ier in order to be able solve complex AI classi4ication problems and also to enhance the robustness of the AI system against adversarial attacks. I present in detail the relevant methodology, experimental setup, results, and analysis, in order to show how Neurosymbolic AI can signi4icantly improve the robustness, accuracy, and interpretability of AI systems. This experimental evaluation highlights the signi4icance of developing Neurosymbolic AI approaches to overcome current de4icits in AI systems, thereby offering a more resilient and comprehensive solution to current AI challenges.

The RLGame is a strategy game of two players that is continuously developed since 2001. It was created in order to conduct research at the field of Reinforcement Learning. In this dissertation we suggest a number of scenarios, where the agents (avatars) that are competing in the RLGame could be studied under the presence of resource constraints. The goal of these scenarios is to investigate whether there are indications of behavioral adaptation when the rules of the game essentially change and therefore could lead to explicit or implicit opponent selection. A number of scenarios are theoretically described and some of them are implemented using Eclipse, a Java integrated development environment. These implementations provided some indications that the existence of some resource constraints might affect the behavior of the synthetic agents and at the same time set the base for future research on the RLGame in order for a more dynamic approach to be implemented and the behavioral adaptation of the agents to be investigated in depth.

Our brain is skilled with the ability to perceive and process multimodal stimuli. This process known as crossmodal perceptual integration, has been in the research spotlight for a long time, providing evidence for the integration of information coming from different modalities. Prior experiments on the field mostly utilized pictures and were limited in the semantic content of a single sound or word. The present study aims to investigate crossmodal perceptual integration in realistic conditions using short movieclips (1500ms) and auditory meaningful three-word sentences in cases of target detection judgments. This study (N=36) is the first to introduce trials without a target that always include target-related information, which was present, either only through vision or audition (incongruent movieclips) or through both (congruent movieclips). For each target condition (present or absent) the movieclips were made up of a combination of 12 videos and 12 sentences, which were repeated in a pseudorandomized order four times for each participant (total trials= 288). The results from the two-way repeated measures ANOVA indicate a similar pattern between the two modalities for semantically incongruent movieclips, with statistically lower accuracy scores in trials where the target was present only in one modality (Maudio=0.647, SDaudio=0.305; Mvisual=0.841, SDvisual=0.235), whereas in target absent trials both showed superior performance (Maudio=0.931, SDaudio=0.038; Mvisual= 0.986, SDvisual=0.018). On the other hand, we observed the opposite pattern for semantically congruent movieclips (Target present trials: Maudiovisual=0.981, SDaudiovisual= 0.036 vs. Target absent trials: Maudiovisual=0.898, SDaudiovisual=0.111). Reaction times were the same for the two modalities (F(2,70)=0.384, p=0.683). In accordance with previous research using images and single words, our results show that when auditory and visual information is congruent, performance is superior and when the target is only present through audio but visual information is incongruent, performance is evidently compromised, and vice versa. Regarding the role of semantics, when the audio sentence included a target-related noun accompanied by a semantically incongruent video, accuracy in judgements was statistically better compared to when it was a verb (tincVerb vs. incNoun=-8.428, p< .001; tconVeb vs. incNoun=-4.256, p< .001). The present results could provide more evidence regarding the role of complexity of semantics, and especially the different role verbs and nouns could play in crossmodal perceptual integration in more realistic situations. Our findings can enrich the content of learning techniques, as well as the design of AI models, by taking advantage of the supporting role of semantic audiovisual information, while taking into consideration the confusion that the complexity in semantic information could cause to perception experience.