Combining learning and symbolic planning for robust goal and plan recognition

• Ano de defesa: 2021
• Autor(a) principal: Amado, Leonardo Rosa
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• 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: Goal Recognition; Plan Recognition; Classical Planning; Auto-Encoders; Machine Learning; Reconhecimento de Objetivos; Reconhecimento de Planos; Planejamento Clássico; Aprendizado de Máquina; CIENCIA DA COMPUTACAO::TEORIA DA COMPUTACAO
• Link de acesso: https://tede2.pucrs.br/tede2/handle/tede/10394

Recent approaches to goal and plan recognition have progressively relaxed the require­ ments about the amount of domain knowledge and available observations, yielding accurate and efficient algorithms. These approaches, however, make two key assumptions about the infor­mation available to the recognizer. First, they assume that there is a domain expert capable of building complete and correct domain knowledge to successfully recognize an agent's goal. Second, even with a complete and correct domain knowledge, most plan recognition approaches are directly affected by the quality of such observations. Such shortcomings can limit the ap­plication of such techniques in real-world applications. While symbolic approaches can provide provable solutions to such problems, learning approaches are adept at dealing with incomplete and noisy data. ln this thesis, we introduce three approaches that improve the performance of goal and plan recognition by combining learning and symbolic planning techniques. First, we use deep unsupervised learning to generate domain theories from data streams (images) and use the resulting domain theories to deal recognize goals in image-based problems. Second, we develop an approach leveraging attention networks to enhance the observation traces of goal recognition problems by predicting missing observations. Third, we combine learning and symbolic planning techniques to compensate for noise and missing observations into new and efficient goal and plan recognition techniques. We show the effectiveness of each technique in a number of domains, ranging from classical domains from planning competitions to image-based domains.