BehavE: Behaviour Understanding through Situation Models for Situation-aware AssistancE

点击翻译按钮获取中文摘要

Situation models allow representing domain knowledge about persons in a structured and consolidated manner. These models are later used for reasoning about the person's behaviour, needs and assistance strategies. Currently, situation models are either built manually or when generated automatically, only a few information sources are used. To address this problem, this project aims at developing a generalised methodology for generating situation models from various heterogenous sources. This methodology will enable the learning of models for different problem domains. More precisely, it will address the following problems: (1) automatically extracting the domain elements and semantics from heterogenous sources; (2) automatically consolidating the heterogenous knowledge into a unified situation model; (3) automatically optimising the learned model based on observed user preferences; (4) automatically maintaining and curating the model over long periods of time so it represents the current situation; (5) developing an evaluation methodology for situation models for real world problems.To achieve that, it will combine existing and novel methods that address different problems of knowledge extraction and model learning from heterogenous sources. They include supervised and unsupervised techniques for semantics extraction and relations discovery; making use of existing structured knowledge to improve the discovered semantics, reinforcement learning techniques for optimising the situation model, as well as various machine learning techniques for maintaining the model and learning the model heuristics. To evaluate the approach, the learned models are applied to existing datasets from the elderly care and healthcare domains and their performance compared to that of handcrafted models. The proposed approach will allow us to reduce the need of expert knowledge and manual development by replacing it with automatically extracted models. If successful, the approach will reduce the time and resources needed for building rich high quality situation models and for developing any system that relies on domain knowledge in order to reason about the solution of a given problem.