Previous: Introduction Up: Issues in Knowledge Level Modelling Next: The Knowledge Level in Practice

The Knowledge Level According to Newell

The knowledge level [28] provides the means to 'rationalise' the behaviour of a system from the standpoint of an external observer. This observer treats the system as a 'black box' but maintains that it acts 'as if' it possesses certain knowledge about the world and uses this knowledge in a perfectly rational way toward reaching its goals. The behaviour of the agent is explained and predicted in terms of the reasons that the agent is assumed to have to take certain actions in order to reach ascribed goals.

In more detail a knowledge level description is based on the following model of the behaviour of an agent [17]:

The agent possesses knowledge

Some of this knowledge constitutes the goals of the agent

The agent has the ability to perform a set of actions

The agent chooses actions according to the principle of rationality:

The agent will select an action to perform next which according to its knowledge leads to the achievement of one of its goals.

The knowledge level, according to Newell, is a level of description above the symbol level. The symbol level provides a means to 'mechanise' a behaviour. At the symbol level a system is described as a mechanism over symbols and representations (structures of symbols). Neither type of description makes a claim about the real nature of the agent. The observer only assumes that the agent acts 'as if' it has such knowledge or mechanisms. The crucial difference is that the symbol level is system oriented, whereas the knowledge level is world oriented. A knowledge level model is in terms of knowledge about the world and the environment in which the agent's problems and actions occur (see Fig. ).

To summarise:

Note that the object of modelling at the knowledge level is not knowledge but behaviour, i.e., observed interaction between an agent and its environment [13]. A knowledge level model is a model of behaviour in terms of knowledge, just like a symbol level model is a model of interaction in terms of symbols and representations. What ties all these models together is the fact that they all model one and the same thing, namely observed interaction. The different models are coherent and consistent to the extend that they model, and to a certain degree predict, the same class of behaviours.

The emphasis on knowledge instead of on representation and implementation issues is the major source of power of the knowledge level idea. It allows one to make meaningful statements about system behaviour without reference to the structures and mechanisms within the agent that realise that behaviour. But consequently it is natural to ask whether a knowledge level model is at all useful to construct systems that behave accordingly. When one writes down (or encodes, for example as a rule) knowledge then it becomes symbol level. The question is, however, whether that encoding can be used to implement the agent. According to the theory it is the observer who creates the knowledge as an annotation of observed behaviour and who uses it at all times to best explain observed and subsequent behaviour. Knowledge should be viewed as an element in the model of behaviour that is constructed by the observer [14]. If it is created through the process of observing interaction then it is not even obvious that knowledge ascribed at some point is also useful for the explanation of subsequent behaviour.

walter@arti