A Loosely Coupled and Distributed Bayesian Framework for Multi-context Recognition in Dynamic Ubiquitous Environments

Today's ubiquitous environments are characterized by smart applications with variable context requirements on the one hand and a dynamic availability of heterogeneous sensors on the other hand. Currently, many existing systems pursue a structured ad-hoc approach with rigid mappings between the applications and the context sources, adversely affecting the performance of these applications when the availability of the context sources drops. Furthermore, such ad-hoc and tightly coupled approaches suffer from reduced flexibility when simultaneously handling multiple smart applications in dynamic environments characterized by a high sensor churn rate. We present a loosely coupled Bayesian-based learning framework that addresses these challenges by allowing dynamic many to many relations between smart applications and context sources with support for recognizing diverse contexts more reliably in the presence of disappearing sensors. Our approach is able to lift these limitations by leveraging the availability of multiple co occurring contexts and their conditional dependencies. On the one hand, the framework exhibits flexibility to dynamically add and remove contexts through autonomic learning of individual contexts appropriate for the spatially distributed ubiquitous infrastructures. On the other hand, it incorporates the advantages of multi-view learning by boot-strapping and fusing multiple heterogeneous context information streams. Our experimental evaluation using a personal assistant application demonstrates the performance and robustness of the proposed framework with significant adaptability and resilience to missing data and partial observability.