Limitations of the unique-attribute representation for a learning system

Two problems faced by many learning systems are that learning can result in an overall increase in system performance time and that overgeneral knowledge can be learned, leading to incorrect system performance. The authors address these problems in an empirical study in the context of the Soar architecture. In this context, they discuss the development of a general unique-attribute representation of annotated models which is used to build a large application with learning capabilities. Besides verifying, on a large application, the previously-reported results regarding the efficiency benefits of a unique-attribute representation, it is shown that the unique-attribute representation can prevent learning overgeneral knowledge. However, the study also reveals the limitations of such a representation. The limitations revealed are that overspecific knowledge can be learned due to the representation; programming within such a representation is more difficult than that within a multi-attribute representation; and proposing operators in parallel, which is essential for flexible problem solving, is problematic