Online System Evaluation and Learning of Data Source Models: a Probabilistic Generative Approach

This paper introduces a method for reasoning about the quality and relevance of data sources supplying inputs to information fusion systems. The emphasis is on probabilistic inference using causal Bayesian networks representing stochastic data generation processes. Modelling patterns that implement explaining away support simultaneous reasoning about (i) hidden states of the domain, (ii) status/quality of individual sources and (iii) the suitability of a specific source type in a given situation. Through such reasoning, a fusion system can estimate the quality and relevance of data sources at runtime which, in turn, is used for automatic adaptation of the source models. This can significantly improve the performance of the overall fusion process. Moreover, the source models feature latent variables representing the status of an individual source and the operational conditions influencing the source, respectively. Consequently, the parameters of such models must be extracted from incomplete training data sets (i.e. there are no observations of the latent states). It is shown that the Expectation Maximization algorithm (EM) in combination with special modelling patterns can cope with this challenge. The effectiveness of the approach is validated with the help of experiments using synthetic data.