The Bayeslet Concept for modular Context Inference

Abstract

In the development and design of ubiquitous computing many challenges are arising. While there is much research done on service management systems and context provisioning, less effort is spent on the methods to actually generate context information - the process of context inference. If we are considering this field of research, we have not only to consider the pure algorithmic problem to infer otherwise unknown information from available data, we also have to target the challenges of large scale systems with millions of users possibly spread across the world and the user’s requirements who is neither willing nor able to wait more than a couple of seconds for his request to be served. In this work we consider shortcomings in today’s context inference systems and analyze requirements for emerging architectures relying on probabilistic algorithms, more precisely static Bayesian networks. We postulate the fragmentation of large networks into smaller so called Bayeslets, that are modular, (un)pluggable, individualisable and easy to process, as they are small and processing can be parallelised. Further on, we propose a formalism to note those Bayeslets in the Bayeslet Language (BalL). Hence, we have a way to easily exchange and deploy Bayeslets and even give application developers a way to provide their own inference rules to the pervasive system.