Learning from Solution Paths

Abstract

Starting with only the legal conditions on a set of operators, a strategy learning system can employ weak methods to search a problem space. Once a solution path has been found, it can be used to assign credit and blame to instances of these operators. If a move leads from a state on the solution path to another state on the solution path, it is labelled as a positive instance of the responsible operator. However, if a move leads from a state on the solution path to a state not on the path, it is marked as a negative instance. These classifications can then be input to a condition-finding mechanism (such as the generalization, discrimination or version space <256> methods); which will determine the heuristically useful conditions under which each operator should be applied.