Pluto: Managing Multistrategy Learning Through Planning

Abstract

Multistrategy learning systems are systems that employ multiple methods to solve learning problems. In many multistrategy systems, either the user or the system selects a single method to use on the current problem. At best, such a selection-type multistrategy learning system can solve the union of the problems solvable by the individual learning methods it contains. Another strategy is to have the user specify a sequence of methods to apply to a particular problem at compile time. Both of these strategies fails to tap the full potential of the learning methods under their control. One way to overcome this limitation is to cast the learning task as a planning problem. By treating learning strategies as operators in a planning problem, several learning strategies can be linked together to form a network bridging the gap between the system’s initial knowledge state, current inputs, and some knowledge goal. Coordinated networks of learning strategies can solve problems beyond the range of any individual learning strategy. We are currently developing a computational model of the planning-to-learn process which we call PLUTO (Planning to Learn Using Transmutation Operators). The current version of PLUTO creates and executes learning plans in the domain of consumer decision making.