12:00, Wed 20 Nov 1996, WeH 7220 Constructing a Learning Strategy under Reasoning Failure Michael T. Cox Michalski formulated a challenge to the machine learning community when he called on researchers to begin to work in the area of multistrategy learning; that is, to explore the space of integrations of learning algorithms, representations, and formalisms in order to build systems that can use different methods and combinations of methods in different learning situations. In this talk, we will examine an attempt to stretch the metaphor of nonlinear planning in the learning-strategy construction task. The general problem is, given a performance failure and a suite of learning algorithms, how can a system choose and sequence calls to the algorithms in order to reduce the likelihood of repeating the failure while, at the same time, avoiding interactions between algorithms. Treating learning as a planning problem, we have specified a collection of learning goals that represent desired changes to the background knowledge and have formalized a series of learning operators in Tate's Task Formalism whose primitive steps represent calls to learning algorithms. We will describe an implemented multistrategy learning system that has tested this metaphor and will show some results that support a deliberate, planful approach to learning. Moreover, we will also show results that indicate, under conditions where learning algorithms interact, non-deliberative learning can lead to worse performance than no learning at all.