NOSS: Declarative Framework for Learning and Evaluating Probabilistic Models of Events in Sensor Networks

Programming a sensor network is difficult, since the programmer has to juggle low-level details of distributed computing with severe resource constraints, in the presence of noisy data and unreliable components. This project focuses on high-level specification of events and activities in sensor networks, since sensor networks are typically deployed for collaborative detection of events and activities. In particular, the project uses a declarative programming framework based on probabilistic logic for high-level specification of events in sensor networks. The probability distributions embedded in the user program are automatically learnt from training examples using standard machine learning techniques. The above approach facilitates high-level specification of sensor network applications, which is automatically translated into low-level distributed code running on individual sensor nodes. The user is thus freed from the burden of worrying about low-level details. The project focuses on the following three goals. The first goal is development of a query engine for efficient distributed evaluation of probabilistic deductive queries in sensor networks. The second goal is to develop techniques for efficient estimation (and distributed re-estimation) of probability distributions embedded in the given program. The third goal is to test the viability of the developed techniques by building two appropriate testbeds. The research project has a significant impact on the ease of programming various sensor network applications. The results of the project are disseminated over the internet at http://www.cs.sunysb.edu/~hgupta/TrainSense.