Actor complexification in TD3 and curriculum learning with structural composition for drone countering

Abstract

The work suggests complexifying actors within the framework of TD3 which involves the usage of different state vectors for actors and critics in order to assure convergence of the algorithm. It also describes a process of aggregating models, separately trained on datasets or in simulation on tasks with increasing difficulty, stitching everything together step by step into a single end-to-end system. It allows utilizing existing algorithms, such as YOLO, in reinforcement learning systems, performing sensor fusion and gradually adding functionality without losing convergence. Assistance providing allows training systems in simulation from hardcoded algorithms that use simplified states. These techniques are demonstrated on a particular task of building an anti-drone system for armored vehicles