An Architecture for Context Prediction


Pervasive Computing is a new area of research with increasing prominence;it is situated at the intersection between human/computer interaction,embedded and distributed systems and networking technology. Its declaredaim is a holistic design of computer systems, which is often describedas the disappearance of computer technology into the periphery ofdaily life. One central aspect of this vision is a partial replacementof explicit, obtrusive interfaces for human/computer interactionthat demand exclusive user attention with implicit ones embeddedinto real-world artifacts that allow intuitive and unobtrusive use.This kind of interaction with computer systems suits human usersbetter, but necessitates an adaption of such systems to the respectivecontext in which they are used. Context is, in this regard, understoodas any information about the current situation of a person, placeor object that is relevant to the user interaction. Context-basedinteraction, which is pursued by the design and implementation ofcontext-sensitive systems, is therefore one of the building blocksof Pervasive Computing. Within the last five years, a number of seminalpublications on the recognition of current context from a combinationof different sensors have been written within this field.This dissertation tackles the next logical step after the recognitionof the current context: the prediction of future contexts. The generalconcept is the prediction of abstract contexts to allow computersystems to proactively prepare for future situations. This kind ofhigh-level context prediction allows an integral consideration ofall ascertainable aspects of context, in contrast to the autonomousprediction of individual aspects like the geographical position ofthe user. It allows to consider patterns and interrelations in theuser behavior which are not apparent at the lower levels of raw sensordata. The present thesis analyzes prerequisites for user-centeredprediction of context and presents an architecture for autonomous,background context recognition and prediction, building upon establishedmethods for data based prediction like the various instances of Markovmodels. Especial attention is turned to implicit user interactionto prevent disruptions of users during their normal tasks and tocontinuous adaption of the developed systems to changed conditions.Another considered aspect is the economical use of resources to allowan integration of context prediction into embedded systems. The developedarchitecture is being implemented in terms of a flexible softwareframework and evaluated with recorded real-world data from everydaysituations. This examination shows that the prediction of abstractcontexts is already possible within certain limits, but that thereis still room for future improvements of the prediction quality.