Advances in imitation learning from observation

• Ano de defesa: 2023
• Autor(a) principal: Santos Júnior, Juarez Monteiro dos
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Pontifícia Universidade Católica do Rio Grande do Sul; Escola Politécnica; Brasil; PUCRS; Programa de Pós-Graduação em Ciência da Computação
• Palavras-chave em Português: Imitation Learning; Behavioral Cloning; Self-Supervised learning; Aprendizado por Imitação; Aprendizado por Observação; Aprendizado Profundo; CIENCIA DA COMPUTACAO::TEORIA DA COMPUTACAO
• Link de acesso: https://tede2.pucrs.br/tede2/handle/tede/11131

Imitation from Observation, a computational technique that instructs agents by observing expert demonstrations, suffers from considerable hurdles such as sub-optimal performance, local minima issues, and ineffective state-space exploration. Although recent strategies leverage unlabeled data to decode information self-supervisedly, persistent challenges remain. This thesis presents four novel methods for imitation learning from observation in response to those challenges. Furthermore, a comprehensive study on the resilience of imitation learning methods is provided to enable a nuanced comprehension of their robustness and performance across various scenarios. The achieved positive outcomes substantiate the merits of the proposed methods. A sampling mechanism is shown to enhance iterative learning cycles, rendering them more balanced. Integrating an exploration mechanism shows potential to surpass expert performance, establishing state-of-the-art results in the field. Moreover, the employment of reinforcement and adversarial learning mechanisms demonstrate their ability to forge more efficient policies, accomplishing good results with fewer samples. The proposed strategies boost performance and efficiency while minimizing the complexity of acquiring expert data