6 Empirical Results on Learning Rate 7 Related Work 4 Intermediate Concept Utilization 5 Intermediate Concept Creation Constructive Induction in Theory Renement

in KBANN (which translates a theory into a neural-net, renes it using backpropagation, and then retranslates the result back into rules) by adding extra hidden units and connections to the initial network; however, this would require predetermining the number and type of intermediate concepts to be created. In addition, unlike EITHER, KBANN is not guaranteed to produce a revised theory that is consistent with the training data. 8 Conclusions In this paper, we have presented constructive induction techniques recently added to the EITHER theory renement system. Intermediate concept utilization employs existing rules in the theory to derive higher-level features for use in induction. Intermediate concept creation employs inverse resolution to introduce new intermediate concepts in order to ll gaps in a theory than span multiple levels. These revisions allow EITHER to make use of imperfect domain theories in the ways typical of previous work in both constructive induction and theory renement. As a result, EITHER is able to handle a wider range of theory imperfections than any other existing theory renement system. References [Danyluk, 1989] A. P. Danyluk. Finding new rules for incomplete theories: Explicit biases for induction with contextual information. than the previous reduction, the process halts. The overall process is halted when no further reduction is possible. Once A has been computed for each case, the reduction operator which produces the greatest syntactic reduction is chosen. In case of ties, intra-construction is chosen since it focuses the reduction on rules having a single consequent. As an example of intermediate concept creation, consider the case in which all of the rules for both liftable and graspable are deleted from cup theory. Given 50 examples, EITHER initially learns the rules (liftable) (has-handle) (weight ?G0009) (< ?G0009 1.1257166) (liftable) (insulating) (width small) (not (has-handle)) (weight ?G0009) (< ?G0009 1.1257166) These rules are then reduced to: (liftable) (intra-0010) (weight ?G0009) (< ?G0009 1.1257166) (intra-0010) (has-handle) (intra-0010) (insulating) (width small) (not (has-handle)) The intermediate concept intra-0010 formed using intra-construction is EITHER's new concept for gras-pable. The extra (not (has-handle)) antecedent on the second rule is a side-eect of translating ID3 decision trees into rules. It does not eect the semantics of the new concept and could be deleted using the sort of rule simplication methods discussed in [Quinlan, 1987]. In order to demonstrate that EITHER's ability to re-ne arbitrarily imperfect domain theories actually increases it's learning rate, this section presents learning curves for …