Abstract
A first order regression algorithm capable of handling real-valued (continuous) variables is introduced and some of its applications are presented. Regressional learning assumes real-valued class and discrete or real-valued variables. The algorithm combines regressional learning with standard ILP concepts, such as first order concept description and background knowledge. A clause is generated by successively refining the initial clause by adding literals of the form A = v for the discrete attributes, A <= v and A >= v for the real-valued attributes, and background knowledge literals to the clause body. The algorithm employs a covering approach (beam search), a heuristic impurity function, and stopping criteria based on local improvement, minimum number of examples, maximum clause length, minimum local improvement, minimum description length, allowed error, and variable depth. An outline of the algorithm and the results of the system's application in some artificial and real-world domains are presented. The real-world domains comprise: modelling of the water behavior in a surge tank, modelling of the workpiece roughness in a steel grinding process and modelling of the operator's behavior during the process of electrical discharge machining. Special emphasis is given to the evaluation of obtained models by domain experts and their comments on the aspects of practical use of the induced knowledge. The results obtained during the knowledge acquisition process show several important guidelines for knowledge acquisition, concerning mainly the process of interaction with domain experts, exposing primarily the importance of comprehensibility of the induced knowledge.