|
|
- Obtaining conclusions. At this point, we have to examine each reasoning step detected when identifying the expert strategic knowledge. If a reasoning step specifies that the system should infer a particular attribute, we need to define how the ES can infer this attribute in all the possible circumstances. So we need to (a), identify the attributes on which this type of inferences may be based; (b) examine whether the inferences of this type have a standard structure (are all these inferences based on the same attributes?); and (c) select an intermediate representation that illustrates how the expert makes and how the ES should make this type of inference.
- Reaction to new information. Some ES have to react immediately, in real time, to the input data they receive or the conclusions reached. In these systems, some reasoning steps will specify that the system should infer the consequences of a new data item or recent conclusion. This knowledge is essential in industrial systems and very important in biological systems. For example, an ES that monitors sensors and is part of a diagnostic and repair/treatment system in a processing plant or in an intensive care unit will not infer a standard set of attributes each time it reads the data from the sensors; it will compare the sensor data with normal measures to detect unusual conditions. When an abnormal condition is detected, it will infer the plant attributes resulting from the abnormal condition, as shown in Figure 5, which represents an organizational chart for high-level control of a simple monitoring application. Note that step 3 is a reasoning step specifying an important basic attribute, but not a specific end attribute. Depending on the condition identified, the system may infer different end attributes. In a reasoning step, such as step 3 in Figure 5, information appears about some basic attributes that fire the inference process. This attribute is referred to as the firing attribute. When analyzing a reasoning step that specifies a firing attribute, it is important to describe how the system should infer the effects of this firing attribute in all possible circumstances. So, we have to (a) identify all the end attributes that can be inferred from the firing attribute; (b) examine which other basic attributes are required in conjunction with the firing attribute; (c) determine whether inferences used by the firing attribute have a standard structure; and (d) select the intermediate representations that illustrate how the ES should infer the effects of the firing attribute.
The intermediate representations described above are just as relevant for tactical knowledge fired when the ES receives information. However, instead of grouping the inferences by the end attribute, they should be grouped by the firing attribute. So, decision tables should be used if the firing attribute is used along with a small number of additional basic attributes to conclude a fixed set of end attributes. Pseudorules are preferable if the system is to make very different types of inferences in different circumstances. It is better to use a decision tree if given characteristics have to be considered in a specified order. Formula definitions should be employed to describe calculations if the attribute is used to calculate other attributes. Finally, a step definition should be created if a procedure to be followed by the ES to infer consequences from the firing attribute needs to be described.
- Inference structure description. If the ES gets several intermediate levels of output conclusions, an inference net may be of assistance in understanding the interdependencies between the attributes of a case. An inference net is a directed graph that shows the overall structure of the inferences within an ES. The net illustrates how the different attributes are inferred and used during the system inference process. Each node in the graph represents an attribute, each arc connecting attributes indicates that the first attribute can be used as a basis for inferring the second. If the ES reasons about a large number of attributes, several inference nets may be needed to illustrate the total system inference structure.
FIGURE 5 Diagram of reaction to new information.
As an inference net provides a perspective of the interdependencies in a conceptual model, this representation is a particularly useful tool during ES extension and refinement. When the definition of an attribute needs to be modified, the inference net shows what other attributes may be affected by the modification.
The analysis of acquired knowledge is concluded by determining factual knowledge. This knowledge contains information known a priori by the ES about the application area, and information that the system will gain about a specific case by executing the task. Here, first the information collected about each general attribute has to be organized within the written definition of the attribute; then the application-relevant attributes have to be classified and the facts about the application area, relating case-independent concepts, processes, or conceptual entities have to be organized. Finally, the interrelations between the concepts, processes, and other conceptual entities identified have to be defined. All these steps are discussed in further detail below.
|