Minimum complexity PDF estimation for correlated data

We investigate the previously overlooked issue in pdf (probability density function) estimation using radial basis functions (RBF), namely the effect of data correlation on the complexity of the RBF network architecture. We propose two simple scalar measures of data correlation. We then introduce a maximum penalized likelihood (MPL) function as a performance criterion for training such pdf estimators where the penalty term is defined in terms of these scalar measures. The proposed MPL criterion favours minimum complexity estimators. The advocated methodology is validated experimentally.