Case-Based Planning to Learn

Learning can be viewed as a problem of planning a series of modiications to memory. We adopt this view of learning and propose the applicability of the case-based planning methodology to the task of planning to learn. We argue that relatively simple, ne-grained primitive inferential operators are needed to support exible planning. We show that it is possible to obtain the beneets of case-based reasoning within a planning to learn framework. A recent view of learning is one in which learning is modeled as an active, deliberative, goal-driven process involving the explicit identiication and pursuit of learning goals 12, 13]. In this view, modiications to memory are treated as planning operations and learning is done by constructing a plan over these operations. The work that has been done in this area has largely focussed on search-based planning algorithms such as non-linear planning; examples of this sort of work include 1, 3, 10, 14]. Furthermore, past work in this area has generally made use of large, complex learning algorithms as the operators for the planning algorithm; for example, Meta-AQUA 1] has a library of strategies such as explanation-based generalization which it combines to form learning plans. We argue, however, that a system capable of a truly broad and exible range of learning needs to be able to construct its own learning algorithms from more basic components. One approach to such basic components can be found in 5] which describes a taxonomy of knowledge transmutations which can be thought of as basic operations over knowledge elements. An example of one of these transmutations is generalize. The generalize transmutation speciies that given a fact about a knowledge element one can infer that information about a more general element which that element is an instance of. Our implementation uses a set of operations which is roughly (but not exactly) a subset of those deened by Michalski; see 6] for further discussion of knowledge planning using transmutation operators. In this paper we are concerned with the algorithms that are used to plan using such operators. Search-based methods for planning are powerful and broadly applicable. However, using very low level primitive operations such as Michalski's knowledge transmutations means that the plans themselves are substantially longer than