Reinforcement learning for efficient UAV-based computer visionColnar, Brin (Avtor) Pejović, Veljko (Mentor) Machidon, Alina - Luminita (Komentor) reinforcement learningunmanned aerial vehicleprecision agriculturecomputer visionWeed control in precision agriculture illustrates the broader challenge of optimizing operational efficiency in dynamic environments - a principle relevant to fields as diverse as financial markets and environmental monitoring. To effectively meet these diverse needs, we have developed a suite of versatile algorithms that select the most appropriate machine learning model for weed recognition in aerial images in real-time, based on context and operational constraints as a UAV flies over an agricultural field. Our algorithms dynamically choose from several pruned versions of the U-net neural network—ranging from 25% to 100% of the original model's capacity—balancing accuracy against resource consumption. Our approach has proven effective, matching or surpassing fixed-model methods. For instance, our upper confidence bandit is able to achieve same Intersection over Union (IoU) metric as 25% pruned network but at a 15% weight reduction, optimizing energy use. Our algorithms show adaptability, as they are able to prioritize different operational needs, such as extending drone flight time or maximizing segmentation accuracy, depending on the situation.20242024-12-20 14:45:01Magistrsko delo/naloga166126sl