On the Limitations of Memory Based Reasoning

Abstract

Memory-Based Reasoning (MBR) represents a radical new departure in AI research. Whereas work in symbolic AI is based on inference and knowledge representation MBR depends on using a large memory of examples as a reasoning base. The MBR methodology is empirical so a typical system does not contain an explicit domain model. This means that MBR systems are quick to set up so the methodology shows considerable promise for knowledge based systems development. Indeed some impressive full scale systems have been demonstrated. In this paper we argue that despite this initial success there are considerable limitations to what can be achieved with MBR. We believe that the absence of a domain model means that MBR will not succeed in complex applications. We illustrate problems in natural language processing and planning that will require access to domain theories in their solution. Our conclusion is that the memory oriented philosophy of MBR has advantages but, for truly intelligent systems, this philosophy is better realised in the CBR paradigm where it can be integrated with a strong domain theory.