MULTISENSOR FUSION UNDER UNKNOWN DISTRIBUTIONS Finite-Sample Performance Guarantees

We consider a multiple sensor system such that for each sensor the outputs are related to the actual feature values according to a certain probability distribution. We present an overview of informational and computational aspects of a fuser that is required to combine the sensor outputs to more accurately predict the feature, when the sensor distributions are unknown but iid measurements are given. Our focus is on methods to compute a fuser with probabilistic guarantees in terms of distribution-free performance bounds based on a finite sample. We first discuss a number of methods based on the empirical risk minimization approach. These methods yield a fuser which is guaranteed, with a high probability, to be close to an optimal fuser (computable only under a complete knowledge of sensor distributions). Then we describe the isolation fusers that are guaranteed to perform at least as good as the best sensor, and the projective fusers that are guaranteed to perform at least as good as the best subset of sensors. Then we consider physical systems wherein the training data consisting of actual physical values is not available. We discuss methods that utilize the physical laws to obtain a suitable fuser under these conditions.