A Parametric POMDP Framework for Efficient Data Acquisition in Error Prone Wireless Sensor Networks

This paper proposes a data acquisition scheme which aims to satisfy probabilistic confidence requirements of the acquired data in an error prone wireless sensor networks (WSNs). Given a statistical model of real-world sensor data and a user's query, the aim of the scheme is to find a sensor selection scheme which best refines the query answer with acceptable confidence. Since most sensor readings are real-valued, we formulate the data acquisition problem as a parametric partially observable Markov decision process (PPOMDP). An existing tool used for solving PPOMDPs, called the fitted value iteration (FVI), is then applied to find a near-optimal sensor selection scheme. Numerical results show that the FVI scheme can achieve near-optimal average long-term rewards, and attain high average confidence levels when compared to other existing algorithms.