Generación de Movimientos Coordinados de Enjambre en Múltiples Drones a través de Algoritmos de Aprendizaje Profundo

Abstract

This degree project proposes a deep learning algorithm based on Q learning that allows a group of agents to represent a swarm movement, specifically leader follower, implementing a repulsion between agents and evasion of fixed obstacles. The learning model includes two methods to reduce the risk of algorithm divergence, the first of which is the inclusion of a memory of experiences for the system and on the other hand the use of a second . The convergence of the Algorithm achieved in less than 6000 episodes was verified with the help of the MATPLOT library to later be implemented in the simulation environment of the Coppelia Sim software. The evaluation of the model implementation system was carried out through 6 experiments, each one of them representing different situations of obstacle avoidance and leader follow-up, demonstrating that the trained model correctly complies with what is expected.