Explanation-Based Generalization for Negation as Failure and Multiple Examples

We report on a combined approach to solve two known problems of traditional Explanation-Based Generalization (EBG) concerning utility and expressiveness. (1) Usually, for each training example a new rule is derived and added separately. Therefore, the overall performance of the associated inference system may degrade if the presented instances are numerous and not representative for the class of applications. (2) Traditional EBG cannot be used for normal domain theories in which negations appear in clause bodies. This lack of expressiveness excludes otherwise interesting applications, e.g. in the area of game playing. As a possible solution for the utility problem (1), it has been proposed earlier [?, ?, ?, ?, ?] to consider multiple examples at the same time and to find a common generalization, thereby compensating for the influence of misleading ones. (2) An extension of EBG for Negation as Failure was described in [?]. However, it is based on Explanation-Based Generalization of Failure (EBGF) [?] which can suffer from an inefficient, redundant representation of rules. This weakness is closely related to a similar one in EBG (1) which is caused by considering training examples independently of each other. The new approach presented here overcomes this drawback by using more than one example also for EBGF. Moreover, we propose a unifying framework for EBG which is able to deal as well with multiple training instances simultaneously as with negation in domain theories.