Institutional Repository
Technical University of Crete
EN  |  EL



My Space

Machine learning and multiagent systems for effective human-machine teaming

Trigkas Nikolaos

Full record

Year 2023
Type of Item Master Thesis
Bibliographic Citation Nikolaos Trigkas, "Machine learning and multiagent systems for effective human-machine teaming", Master Thesis, School of Electrical and Computer Engineering, Technical University of Crete, Chania, Greece, 2023
Appears in Collections


Τhis MSc thesis proposes innovative machine learning and multiagent systems algorithms, for the effective collaboration of humans and machines. More specifically, it puts forward a novel AI-based algorithmic ``toolkit’’ in order to enable teams of Unmanned Aerial Vehicles (UAVs) to assist in the victims’ localization and rescue, in case of post-avalanche events. Our thesis consists of two main research axes that gave rise to respective elements within the aforementioned algorithmic toolkit: first, the processing and analysis of post-avalanche scenery images (in order for the UAVs to be able to recognize the victims), via the employment and enhancement of several well-known image processing techniques; and second, the creation of a novel coalition formation framework that facilitates the UAVs' cooperation and coordination.In some detail, for the victims’ recognition task itself, we employ three image recognition algorithms, putting forward object and edge detection techniques used for the first time for the analysis of post-avalanche Search & Rescue operations images. More specifically, we apply (a) a Colour Desaturation Method in which we manage image information to be reduced by using color filtering; (b) a novel version of the well-known Sobel edge detection algorithm, which enhances the tracked edges of the background and the desired objects in it; and (c) the ``Faster R-CNN’’ object detection method, offering state-of the-art region selection and image segmentation. All three algorithms are tested in real-time simulations, with the best of which in combinational effectiveness (i.e. the Faster R-CNN) being used during later stages of our research.At the same time, the second axis of our thesis work involves putting forward a novel coalition formation framework consisting of multiple components. First, a proposed initial UAVs placement algorithm (adapting a known Brain-Storming Optimization algorithm to our setting); second, a coalition structure generation protocol that allows for the "online" calculation of coalition values (representing the value of each potential teaming-up configuration of the UAVs at hand) that will eventually guide the rescue effort; and third, a simple but effective opinion aggregation protocol, that can be used to prioritize rescue operations in the event of "ambiguous" findings. The combination of the above modules aims to maximize the number of victims that can be tracked and rescued based on cooperative discoveries by the UAVs, and to completing this process in the shortest possible amount of time. Our experimental evaluation verifies the applicability and effectiveness of our framework and its individual components, in a variety of different scenario simulations. Finally, our thesis work led to a research paper entitled "Coalitions of UAVs for Victims Localization in Post-Avalanche Events Using Advanced Image Processing Techniques & Algorithms", published after peer-review in the proceedings of the (international) 12th Hellenic Conference of Artificial Intelligence (SETN 2022).

Available Files