1991 Volume 6 Issue 1 Pages 72-83
The augmented EBL is a framework for knowledge refinement based on generalization from examples. It is different from conventional EBL that plural examples are considered simultaneously. It has been formalized on logic program, and relationships between generalization space of plural examples and operationality criteria have been discussed. In this paper, we forcus on a class of problem solving that have serially decomposable subgoals, and propose an augmented EBL learner that acquires a problem solving macrotable from examples. The learning problem is formalized on logic program. A domain theory consists of operator definitions which is the minimal knowledge for problem solving. A solution order is a sequence of generalizations of the goal. A macro table is a logic program, which is ordered reversally to the solution order. Given operator sequences of macros and their row positions in the macro table, it is shown easy to determine the instantiation of macros, that is precondition and conclusion, and the solution order. Thus it is substantial to determine operator sequences and their row positions in order to learn a macro table. We propose an augmented EBL learner that uses membership queries in addition with randomly given positive examples. Membership queries are used through two concepts of a decomposition and a serialization table. An example is decomposed into a composition of macros. The serialization table enumerates all feasible orderings of 2 macros according to results of membership queries against their compositions. The usefulness of the learner is shown by applying to 8-puzzle. The serialization table is shown complete, then the learner acquires correct macro table. The complexity of the learner is polynomial in the size of macro table when there exist no composite macros, that is decomposable into other macros. Current implementation generates exponentially many membership queries of the size of composite macros if they exist. It is a further issue to resolve this difficulty.