Previous: The Knowledge Level in Practice Up: Issues in Knowledge Level Modelling Next: Varieties of KL-Models

Two Step Rationality

How can the above be reconciled with Newell's notion of knowledge level? The principle of rationality (Sect. ) relates knowledge to behaviour in a single step [28]. It serves as a global interpreter for the knowledge of the agent in order to understand observed behaviour or to predict future behaviour. Two step rationality [42] views this process as consisting of two steps: configuration and application (Fig. ):

In the first step, called 'configuration', knowledge is configured into a model of the problem-situation that the agent is faced with (called the 'task instance'). In this step the agent is thought of as imposing a structure on selected pieces of its knowledge, and assigning roles to these in its reasoning. This structure is what we have referred to as the KL-model. It captures the way in which the agent frames a task instance. The ability to appropriately configure a KL-model is probably at the heart of excellence in problem solving and rational behaviour. The goal of configuration is to enable the practical application of the principle of rationality within the boundaries of the KL-model. This 'application' is the second step in two step rationality. To the extent that the task model is adequately configured knowledge can be practically applied and leads to satisfactory goal directed behaviour.

What makes every task instance different from the other one are the specific circumstances in which it is to be solved. These circumstances may be related to the characteristics of the environment (critical, static or highly dynamic, size and complexity) and the agent's interaction with it (accessibility and reliability), or may impose additional restrictions on the task (quality or specificity of solution, limited resources). These aspects of a task instance are called the 'task features'. An important class of task features follow from the epistemological problems (i.e. problems with the knowledge) and the pragmatic problems (i.e. problems with the pragmatics of using the knowledge) that an agent is faced with [35]. Epistemological problems arise because we are dealing with models of a real and open-ended world. Every model is an approximation, not only because it is unprecise but, more fundamentally, because it introduces assumptions about the world. Problems of pragmatics are forced upon us by the reality of the task environment. They follow from the practical limitations of the complete system of the agent and its environment, for example the impossibility to make sufficiently precise measurements or to exhaustively explore a range of possible malfunctions. Steels argues that real expert knowledge (for example a situation specific way of dealing with an unknown factor) is aimed exactly at satisficing performance in spite of these problems.

Task features play a fundamental role in configuring the KL-model. The KL-model must be such that the behaviour resulting from the application of the principle of rationality within the boundaries of that model is also practical, i.e., complies with the task features of the task instance. The decomposition of the principle of rationality reflects the fact that there is more to rationality than just being logical. The task features of a task instance determine what behaviour is practical, i.e., feasible and appropriate within the actual agent-task-environment combination. Once a KL-model has been configured to comply with these practical constraints the principle of rationality can freely explore possibilities. The pragmatic concerns are thus taken care of in configuration, the rational concerns by the application of the principle of rationality within the KL-model [42].

walter@arti