Control of Perimeter Surveillance Wireless Sensor Networks via Partially Observable Marcov Decision Process

This paper presents a novel approach to controlling large wireless sensor networks capable of optimizing multiple conflicting performance criterion. The example of battery power usage versus target tracking error is formulated here, though the technique can be extended to assessment, false alarm reduction, etc. Modeling a perimeter security system as a partially observable Markov decision process, an intruder's behaviors are probabilistically estimated several steps into the future (look-ahead) thus allowing the system to make the best decisions for overall benefit (non-myopic). In this example sensor activation is the control input. Further, particle filtering is employed to improve location estimates of multiple targets from noisy sensor data. Performance of the algorithm is demonstrated on a high fidelity simulator called UMBRA. Also a 100 node wireless sensor network has been constructed for algorithm validation