University of Thessaly (UTH)
Bio-inspired algorithms for data analysis and decision-making
Abstract
dc:descriptionThis doctoral dissertation investigates bio-inspired algorithms and models for decision-making and data analysis. In view of scientific challenges for developing innovative solutions with high social impact, it explores applications in the context of image and signal analysis, including interpretable image classification, emotion recognition, depression severity estimation, risk transmission of infectious diseases, assistive navigation of visually impaired individuals, and virtual coaching for physical exercise. In this context, it focuses on the development of interpretable decision-making frameworks and models, based on fuzzy logic, and their optimization using bio-inspired algorithms. Optimization algorithms constitute a powerful tool and are used in many decision-making tasks. Although a plethora of such algorithms has been proposed there are still directions for improvement. For example, many of these algorithms have limitations like premature convergence, the possibility of being trapped into local optima, and limited diversity in generating optimal solutions. A recent bio-inspired algorithm is Brain Storm Optimization (BSO). The BSO algorithm mimics the human brainstorming process for efficient problem solving. Solutions are modeled as individuals, each representing an idea generated during the brainstorming process. These individuals belong to a larger population, which is clustered and iteratively updated by selecting a subset of individuals. Specifically, during the clustering process, all individuals are divided into several groups by applying the k-means algorithm. A new individual is then created, using cluster centers or individuals of one or two clusters. The best individual is selected and recorded as the final solution to the given problem. In BSO and most of its current modifications, the termination of the algorithm usually depends on a predefined upper bound of algorithm iterations, whereas its convergence speed needs improvement. To deal with the above limitations, in this dissertation, two novel modifications of BSO are proposed; firstly, the Determinative BSO (DBSO) algorithm is proposed, which is capable of automatically converging to optimal solutions. Specifically, a key component of the algorithm is a clustering component it includes, which starts with a relatively large number of clusters corresponding to the possible ideas and whenever clusters are identified, the clusters are merged, considering the criteria mentioned above. After the number of clusters has been reduced and there are not any other different enough possible solutions, the algorithm terminates. Another algorithm, named Stepladder Determinative BSO (S-DBSO) has been created based on a brainstorming technique for decision making, proposed in the field of psychology, named stepladder technique; this decision-making tool aims to the equal participation of all members during the brainstorming process, in a way that all of them can express their thoughts. The results of the experiments using well-known benchmark functions, and the scalability investigation performed, confirm that S-DBSO has an optimized directed search ability towards the best solutions, compared to the baseline algorithm as well as other recently proposed state-of-the-art algorithms. Interpretable machine learning (ML) models have received remarkable attention in terms of their capability to provide understandable explanations of their predictions evoked from complex systems. In this dissertation, a novel generalized framework is developed aiming to create interpretable auto-constructed Fuzzy Cognitive Maps (FCMs). Structured knowledge in the form of graphs, including FCMs, provides many benefits, including high understandability and knowledge representation accuracy. However, the manual development of the graphs requires the intervention and participation of experts with specific knowledge. Thus, despite their simple structure and ease of use, current FCMs also have some limitations, as in some cases, no expert may be available to contribute to the definition of an FCM. Furthermore, in complex systems, the large number of concepts and their interconnections that may include make manual model design difficult. Considering the above, the introduced FCM framework embeds a mechanism for automatic determination of the FCM structure from the datasets used in each experiment. This is achieved by considering a training dataset as a source of knowledge to automatically detect and define the cause-and-effect relationships between the concepts of a given problem. The introduced framework has been effectively applied to multi-label image classification as well as to biomedical image analysis, aiming to address the challenge of automated lesion risk assessment based on wireless capsule endoscopy (WCE) images, and pneumonia diagnosis using X-rays. Experiments using publicly available datasets demonstrate the effectiveness of the proposed framework and models and confirm that they can successfully extend the application of baseline FCMs for image analysis. A contribution of this dissertation is the development of FCMs for signal analysis, on a novel application considering electroencephalogram (EEG) signals. In particular, a novel framework for automatically creating FCMs is proposed, and applied for the estimation of depression severity of adolescents. The introduced framework uses features extracted from a model that detects the cause-and-effect relationships using fuzzy logic, between brain activity and depression, in a more intuitive way. Furthermore, an auto-constructed FCM model, using EEG signals, has been developed aiming to perform interpretable emotion recognition. The proposed approaches are simple to implement and earn the trust of the user, while providing interpretable results. Another research direction investigated in this dissertation has led to the development of a novel FCM model, aiming to deal with the challenge of assistive navigation of visually impaired individuals. The proposed model is based on DBSO, which is used to automatically adapt the weights of the FCM structure. This application ensures the avoidance of static obstacles in an unknown environment, while considering the output of an obstacle detection system based on a depth camera. The simulation results show that the proposed model can effectively assist the users to avoid static obstacles and safely reach a desired destination. A novel user-centered virtual coaching (VC) framework based on FCMs is also introduced in this work for physical exercise. Τhe introduced framework, an FCM plays the role of a virtual coach as it configures suitable paths for users to navigate, based on their desired exercise intensity, as well as their physiological measurements obtained from wearable sensors, during their exercise. The contributions of the proposed framework include the following: i) a novel VC FCM that automatically determines its structure through a supervised process; ii) a dynamic personalized adaptation mechanism enabling the FCM graph to change its parameters based on the evoked physiological measurements and adapt to any desired path the users want to navigate, iii) an inherent interpretation mechanism capable of explaining the decisions of the FCM to the users, thus gaining their trust. Another contribution of this dissertation is the development of a novel framework based on intuitionistic FCMs (iFCMs) for interpretable image classification. . Intuitionistic FCMs (iFCMs) have been proposed as an extension of FCMs with the aim of offering a natural mechanism for evaluating the quality of their output through the estimation of hesitancy, a concept resembling human hesitation in decision making. In the context of image classification, hesitancy is considered as a degree of unconfidence with which an image is categorized to a class. Specifically, this framework includes a CNN-based feature extraction process that focuses on the most informative image regions. It also presents a methodology for constructing data-driven IFSs and a learning algorithm that automatically defines the graph interconnections based on these IFSs. In addition, an interpretable classification approach is proposed, using the iFCM structure to explain the classification results based on local information derived from image regions. The various methodologies introduced in this dissertation outperform the respective state-of-the-art methods, while they are able to cope with various challenges that have been documented in the literature for each domain of application. The research landscape explored is broad and its contributions are expected to have significant scientific and societal impact, opening new perspectives for further future research and the advancement of science.
Degree
thesis:*- Grantor dc:publisher
- University of Thessaly (UTH)
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Sovatzidi, Georgia
- Σοβατζίδη, Γεωργία
Subjects
dc:subject × 20- Τεχνητή νοημοσύνη (ΤΝ)
- Ανάλυση δεδομένων
- Λήψη αποφάσεων
- Αλγόριθμοι βελτιστοποίησης
- Ασαφή σύνολα και ασαφή λογική
- Ερμηνεύσιμη μηχανική μάθηση
- Επεξεργασία σήματος
- Artificial Intelligence (AI)
- Data analysis
- Decision making
- Optimization algorithms
- Fuzzy sets and fuzzy logic
- Explainable machine learning
- Signal analysis
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Τεχνητή νοημοσύνη
- Computer and Information Sciences
- Natural Sciences
- Artificial Intelligence
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/59280