A new approach to rule induction and pruning

Abstract

The authors regard the rule induction problem as a search in the lattice of possible rules. Two kinds of search directions (top=generalization and down=specialization) and therefore two different kinds of algorithms exist. They propose the algorithms RELAX and its heuristical modification H-RELAX, which both use generalization as the search direction and therefore can be used for rule induction and rule pruning. They produce a minimal set of rules or a set of rules which allow object recognition with minimal costs. Both algorithms also allow an incremental and knowledge-based search process and can be combined with other algorithms.<>