Active Decomposition and Sensing in Networks of Distributed Chemical Sensors

Active sensing enables a sensor to optimize its tunings on-the-fly based on information obtained from previous measurements. When applied to networks of distributed sensors, however, active sensing becomes computationally impractical due to the combinatorial number of sensing configurations. To address this problem, we present an active decomposition and sensing (ADS) method that combines the advantages of classifier decomposition with those of active sensing. Namely, we use class posteriors to decompose the problem across the sensors in the network. Each sensor then applies active sensing to select the next tuning to solve its specific subproblem. As a result, the method scales linearly (rather than combinatorially) with the number of sensors. We validate ADS on a database of infrared absorption spectroscopy containing 50 chemicals. Our results show that active decomposition improves classification performance and reduces sensing costs when compared to using active sensing only at the node level.