Plan recognition and failure prediction for ambient assisted living

• Ano de defesa: 2017
• Autor(a) principal: Farias, Giovani Parente
• 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: Reconhecimento de atividade; Reconhecimento de plano; Reconhecimento de múltiplos objetivos; Predição de falha; Activity recognition; Plan recognition; Multiple-goal recognition; Failure prediction; CIENCIA DA COMPUTACAO::TEORIA DA COMPUTACAO
• Link de acesso: http://tede2.pucrs.br/tede2/handle/tede/9142

The process of inferring agent’s plans/goals from their observed actions is known as plan recognition. Predicting human intentions is one of the ultimate goals of Artificial Intelligence; plan recognition contributes to this goal by analysing how low-level observations about agents and environment can be associated with a high-level plan description. Most approaches to plan recognition, in realistic environments, are based on manually constructed rules, where the knowledge base is represented as a plan library for recognising activities and plans. Besides, these approaches do not usually have the ability to incorporate complex temporal dependencies, and they take the unrealistic assumption that an agent carries out only one activity at a time and the sequence of actions is all coherently executed towards a single goal. Moreover, the incomplete knowledge about the agent’s behaviour and the similarity among several plan execution generate multiple hypotheses about the agent’s plan(s) that are consistent with the observations. This work addresses the problems of recognising multiple plans in realistic environments, learning activity duration, and detecting anomalies in plan execution. We deal with problems related to disambiguation of multiple hypotheses and detecting anomalies in plan sequence by exploiting both the inherent hierarchical organisation of activities and their expected time and duration, developing an efficient algorithm to filter the hypotheses by applying temporal and path length constraints. We present a number of experimental results showing that, besides addressing those limitations of traditional plan recognition algorithms, our filtering approach can significantly improve the accuracy of the underlying plan recognition algorithm. The experiments include a number of synthetically generated plan libraries as well as plan libraries and observations obtained from a real-world dataset useful in the context of ambient assisted living.