A NEW "IMPLICIT" PARAMETER ESTIMATION FOR CONDITIONAL GAUSSIAN BAYESIAN NETWORKS

Among existing Bayesian network (BN) parametrizations, conditional Gaussian are able to deal with discrete and continuous variables. Bayesian estimation of conditional Gaussian parameter needs to define several a priori parameters which are not easily understandable or interpretable for users. The approach we propose here is free from this priors definition. We use the Implicit estimation method which offers a substantial computational advantage for learning from observations without prior knowledge and thus provides a good alternative to Bayesian estimation when priors are missing. We illustrate the interest of such estimation method by first giving the Bayesian Expectation A Posteriori estimator (EAP) for conditional Gaussian parameters. We then describe the Implicit estimator for the same parameters. One experimental study is proposed in order to compare both approaches.