A comparison of four tests for attribute dependency in the LEM and LERS systems for learning from examples

Abstract

The paper discusses two programs for learning from examples, LEM (Learning from Examples Module) and LERS (Learning from Examples based on Rough Sets). A few versions of both programs are implemented in Franz Lisp and are running on VAX 11/780. Both programs' main task is to automate knowledge acquisition for expert systems. Hence, they produce rules in the minimal discriminant form. The main problem addressed in the paper is the selection of the best mechanism for determining coverings, the minimal sets of relevant attributes. Four different methods, based on indiscernibility relation, partition, characteristic set and lower boundary are compared. Both theoretical analysis and experimental results of multiple running of many sets of examples, with variable number of examples and with variable number of attributes are taken into account. As a result the partition method is determined to be the most efficient way to compute coverings.