How it works:
- I. Off-line learning estimation of the statistical model from original census data
- II. on-line computation of conditional histograms from the statistical model
Phase Ia) - Simple learning
Original census data are complete or missing values are considered as specific. The estimation of statistical model parameters is done by standard EM algorithm.
Phase Ib) - Learning from incomplete data
Statistical model is first used to estimate missing values.
When some records are incomplete, the estimation is divided into several steps. First modified EM agorithm is used to get temporary estimation of multivariete distribution, which serves then as a knowledge base for filling in the missing values. The completed data are then used as an input for second learning step, identical to the simple case mentioned above.
Phase II) - Interactive (online) computing
Model is used as a knowledge base for our propabilistic expert system (PES). No need for original data.