The Knowledge Level Hypothesis: Our point of view

The knowledge level hypothesis was introduced by Allen Newell in 1982 (Newell, 1982). Allen Newell formulated this hypothesis in a tentative to solve the confusions in the use of the « knowledge » and « representation » terms. He defined a representation as a symbolic system which codes a body of knowledge. His hypothesis advocates the existence of knowledge independently of its representation [ibid., p. 99]. At this level an agent is described as having a set of objectives, a set of possible actions, and a body of knowledge. His behaviour is predicted by the principle of rationality [ibid., p. 102]. Knowledge is characterised as a notion of competence, being a potential to generate actions. Knowledge is also completely independent of representation and implementation questions.
Clancey was the first to show the relevance of modelling at the knowledge level for the development of knowledge-based systems (e.g., (Clancey, 1982)). Another paper defending knowledge level modelling is (Schreiber, Akkermans and Wielinga, 1989). Following research worked out this point of view. In some recent approaches of knowledge engineering, the idea of knowledge level is a recurrent theme because the separation of the knowledge-oriented aspects from questions of symbolic encoding and computation is viewed as an important principle for structuring and simplifying the management of knowledge-based system projects.
However, the knowledge level idea is rarely applied in its original form. For example in works on KADS (Schreiber, Wielinga and Breuker, 1991) and CommonKADS (Wielinga, Van de Velde, Schreiber and Akkermans, 1993), knowledge is structured: we have, for example, domain knowledge organised around an ontology, task knowledge organised in a hierarchy, and method knowledge that implements tasks. In those approaches, the structure given to the knowledge is as important as the knowledge itself. Indeed, if the structure does not correspond to the problem to solve, the system will not be able to solve it, even if it has sufficient knowledge. This is opposed to the original idea of knowledge level where knowledge was not structured. Therefore, these approaches are not truly at the knowledge level but mix it with compiled knowledge on how the knowledge is structured and how it will be used. They are nevertheless useful, because they restrain the principle of rationality (which is an imprecise concept not really made explicit by Newell) to more applicable principles that depend on the structure (such as generate-and-test, etc).
To summarise, we can say that the original knowledge level principle of rationality has been divided in a « two step rationality » (Van de Velde, 1993). A first step structures the knowledge to what is called in the modelling methods the «knowledge level model». Then a second step fills this model with knowledge. The result is the specification of a system that will solve a problem in the framework of the knowledge level model. Given this two step rationality, we think that the most importance phase for competence is the first one. Indeed:
  1. The first step defines the framework in which solutions can be searched and thus defines the limits of the system. It really defines the system competence. It is surprising to realise that the suppositions resulting from this structuring are rarely made explicit and are not made available to the system itself.
  2. The second step is a much more mechanical phase where the principle of rationality is replaced by a more logical principle. This principle while more amenable to computation is insufficient to account for competence (Blot, Durand, Jézéquel, Malandain and Moulin, 1995).
To conclude on Newell's knowledge level hypothesis, we think that only a part of it has been exploited in the « knowledge level approaches ». This part focuses on making explicit the knowledge used by filling a « knowledge structure ». While being necessary, this step is not sufficient to account for competence. Therefore, we believe that truly competent and co-operative systems should posses knowledge on the first step.

Back Back to the main page