Previous: The Knowledge Level According to Newell Up: Issues in Knowledge Level Modelling Next: Two Step Rationality

The Knowledge Level in Practice

The original aim of the knowledge level was to clear up confusion concerning the usage of the terms 'knowledge' and 'representation'. The idea immediately resonated with ongoing research toward understanding and building knowledge systems from a knowledge content (epistemological) perspective [7][38][10]. Clancey's model of heuristic classification [11] illustrated the power and scope of competence models that make explicit the kinds of knowledge embodied in a system and their roles in an overall pattern of reasoning. Here are only some of the most representative approaches in knowledge engineering that in one form or another use the knowledge level notion:

They take the knowledge level as a source of inspiration or as a way to explain what they are doing. If one looks more closely then one finds striking differences with the original knowledge level descriptions. To distinguish these models from the knowledge level descriptions a la Newell I use the capitalised "Knowledge Level model", or KL-model. The most obvious difference between a knowledge level model and a KL-model is that the latter has lots of structure. As a consequence a KL-model implies other than 'why' aspects of a system's behaviour. For example one finds notions of 'task' and 'method' that seem to relate more to 'what' and 'how' models of behaviour.

So knowledge level models and KL-models are different but related things (Fig. ):

A KL-model is a structure that is imposed on knowledge when it is being put to use in a class of problem situations.

A KL-model describes a structure on knowledge. Like the knowledge itself, this structure is ascribed by an observer. It is a use-specific window on the knowledge level, which in turn is a model of behaviour. For example, the model of heuristic classification is visible in a pattern of inferences that contribute to abstract, heuristic match and refinement steps in reasoning [11]. This well known example also illustrates what is meant by structure, namely role-limitation [26]: the role in reasoning of parts of the knowledge is being specialised. In addition, these parts of the knowledge are being organised, for example into hierarchies or causal networks. What counts in a KL-model, however, is not only the knowledge that the agent seems to be using but, more importantly, the structure within which this knowledge is being used for achieving goals. This structure becomes visible through the process of reasoning.

KL-models are models of specialised intelligence. Newell banned all structure from the knowledge level to account for the fully rational exploitation of an agent's knowledge in a model of general intelligence. The KL-model, on the other hand, is characterised by a deliberate restriction of flexibility. This restriction allows the agent to reach a limited range of goals in a limited range of situations, but to do so in a highly adapted (expert) way. They are expert in a specialised task. Intelligence in this context is not so much related to the capability of producing isolated episodes of rationalisable behaviour but rather to the re-occurrence of a structure of such a rationalisation over a range of behaviours.

walter@arti