Context inference is necessary in ubiquitous computing to provide information about contextual information which is not directly measurable from sensors or obtained from other information sources. Server based, central inference would not scale due to the expected amount of context requests. Mobile, distributed context inference faces problems because of the high computational complexity of inference mechanisms. Bayesian inference techniques are particularly well suited, as they allow for more flexible modelling of situations than propositional logic, are always decidable as opposed to higher order logics, are intelligible to humans as opposed to neural networks and allow for uncertain or missing information. As inference in them however is NP-hard, methods have to be introduced to fit them to the requirements of ubiquitous computing and mobile, resource constrained devices.To this end, this work proposes to divide Bayesian networks for context inference into modules, called Bayeslets. Bayeslets can be composed among each other to fulfil an inference request via interface nodes about which additional assumptions are made: Considering input nodes as observed, more efficient inference methods can possibly be applied and by defining explicit output nodes for connection, a relevancy based dynamic composition of Bayeslets can be realised, so the evaluated number of Bayeslets always stays at a minimum. The inference time of Bayeslets can be further reduced by adapting edges and value ranges to the user’s personal requirements and the current situation. The application of these concepts is shown in general examples of high level context used in the user’s smart space, in his work environment, as well as in road traffic. Experimental results show that this process results in a significant reduction of the inference load. The Bayeslets for location and human motion related activity are of particular importance for context awareness and therefore considered and evaluated in detail.The set of tools proposed in this thesis allows to apply a fully Bayesian approach to context inference, fulfilling the requirements of ubiquitous computing and mobile, resource constrained devices.weiterlesen