Algoritmos de aprendizagem por reforço aplicados à navegação autônoma de robôs

• Autor(a) principal: Costa, Emanoel Lucas Rodrigues
• Data de Publicação: 2017
• Tipo de documento: Trabalho de conclusão de curso
• Idioma: por
• Título da fonte: Repositório Digital da Universidade Federal Rural do Semi-Árido (RDU)
• Texto Completo: https://repositorio.ufersa.edu.br/handle/prefix/6262
• topic: Aprendizagem por Reforço; Navegação Autônoma; Política; Q-learning. Sarsa; Reinforcement Learning; Autonomous Navigation; Policy; Q-learning. Sarsa; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA E TECNOLOGIA

Reinforcement Learning (RL) has proven to be a very efficient method when solving problems where it is necessary to learn a behavior from an unknown environment. There are several applications of this method in the most different areas of research, one of these applications has attracted the attention of researchers in the area of computing, which is the ability to create autonomous agents using reinforcement learning. There are several RL algorithms that can be applied to learning tasks, in order to determine an optimal policy for an agent. This work proposes to perform a study of reinforcement learning algorithms applied to problems of autonomous navigation, for this was done a bibliographic survey on the subject and implemented two algorithms quite popular in literature, Q-learning and Sarsa, seeking to analyze them as to the ability to converge to an optimal policy and its performances in the application in an autonomous navigation model, that seeks to learn to avoid obstacles and reach a predetermined goal. The SCILAB software was used to simulate the developed environment and implementation of the RL algorithms, the results obtained in the tests of both algorithms were analyzed regarding the autonomous learning ability of the agent in the search of the determined objective, where changes were made in the parameters of the learning equations and compared the results in order to obtain the best performance. The algorithms presented different results for each variation of the parameters in the accomplishment of the same task in the developed environment, with highlight to the Sarsa algorithm, which was more sensitive to the variations. So, the performance for a reinforcement learning application depends on both, the algorithms used and the learning parameters used.