Assumed density filtering for learning Gaussian process models

In this paper, we consider the probabilistic modeling of censored sensor readings. Specifically, we first model the sensor observations using Gaussian process framework and develop two computational techniques - one based on Assumed Density Filtering (ADF) and the other based on Monte-Carlo method, for estimating the parameters of the approximate posterior density of the sensor observations. We compare their performances using a simulated sensor network example and show that the ADF-based technique is much faster than the Monte-Carlo-based technique. Further, we also show that our approach performs better than the standard Gaussian process regression technique which simply discards the information from sensors that fail to detect the source phenomena.