Rule extraction by successive regularization

Proposes an approach to rule extraction by successive regularization which generates a small number of dominant rules in an earlier stage and less dominant rules or exceptions at later stages. It is not only computationally robust but also advantageous from a viewpoint of human understanding. Humans tend to interpret data as a small number of dominant rules and their exceptions, instead of a large number of rules. This hierarchical structure of rules and their exceptions is much easier to understand than a non-hierarchical set of rules.