Wireless Sensor Networks are tiny devices equipped with sensors and wireless communication. These devices observe environments and communicatie about these observations. Machine Learning techniques are of interest for Wireless Sensor Network applications since they can reduce the amount of needed communication and as a result the amount of energy that is used. The application of Machine Learning for Wireless Sensor Networks is not straightforward. For example, distributed computations, energy constraints and memory limitations make the careful selection of algorithms and the application architecture critical. Most Machine Learning research was not conducted with these aspects in mind. Feed Forward Neural Networks, for example, have limited options for distributed execution, while the accuracy of Decision Trees is very sensitive to the failure of sensor nodes. This thesis introduces a new classification algorithm, named QUEST: QUantile Estimation after Supervised Training. QUEST is based on the Naive Bayes classifier and was designed to be suitable for Wireless Sensor Networks. Properties of interest that are inherited from regular Naive Bayes are flexibility with regard to distributed implementation and robustness with regard to hardware failure of individual components. A new property of QUEST is that it eliminates the need for online supervised learning by using unsupervised learning techniques to adapt itself to new environments. As such, QUEST enables the efficient deployment of Wireless Sensor Networks and reduces the manual maintenance required in case of battery depletion.
|Qualification||Doctor of Philosophy|
|Award date||24 Feb 2017|
|Place of Publication||Enschede|
|Publication status||Published - 24 Feb 2017|