Transfer of Training in Procedural Learning A Matter of Conjectures and Refutations?

Abstract

Heuristics learning is defined as the process of discovering rules for how to apply problem solving operators. An analysis of transfer of training with respect to problem solving heuristics results in two transfer mechanisms; one based on the interplay between conjectures and refutations, and one based on the partitioning of a goal into independently realizable parts or subgoals. First, rules called proposers produce suggestions about which operator(s) should be considered in the current situation, while other rules — called censors — refute those suggestions which previous experience has shown to be bad. A system which can learn proposers and censors can grow by the successive attenuation of the conditions on the proposers and through the successive addition of censors. In the long run, such a system will apply each problem solving operator when and only when it is appropriate. Second, subgoaling rules encode knowledge of which parts of a goal description can be attained separately. Such rules can be learned by noticing the successive transformations of the current state during problem solving; each step which makes the current state more similar to the goal state defines a potential future subgoal. Subgoaling rules learned while solving one task can facilitate the solution to another task, if the tasks share at least one subgoal. A computer program is written on the basis of this theory, and shown to be able to transfer within a simple task domain.