Systems, Tasks and Adaptation Knowledge: Revealing Some Revealing Dependencies

This paper shows that the use of adaptation knowledge in CBR systems is heavily dependent on certain task and system constraints. Furthermore, the type of adaptation knowledge used in particular systems performing specific tasks is quite regular and predictable. These conclusions are reached by reviewing forty-two CBR systems in the literature and classifying them according to specified taxonomies. Three taxonomies are produced in the course of the paper. First, an adaptation-relevant taxonomy of CBR systems is outlined using four dimensions: whether adaptation is performed or not, whether single or multiple cases are used, whether case solutions are atomic or compound, whether parts of a compound solution interact with one another or not. Second, a taxonomy of tasks is produced based on a similar taxonomy proposed by Clancey (Clancey 1985). Third, a taxonomy of adaptation knowledge that distinguishes target-elaboration, role-substitution, subgoaling and goal-interaction operators is described. We then show how different systems cluster with respect to interactions between these three taxonomies. The conclusions of this paper could be applied as guidelines for system designers or support the comparative evaluation of systems.