Inference Structure for P&R

Propose&Revise (P&R) is a problem-solving method to accomplish a configuration task, which constructs a design by proposing a value for one parameter of a system at a time and checking to see whether each parameter satisfies all constraints on it. P&R was originally implemented by Salt, a knowledge-acquisition tool that generates P&R systems [Marcus, 1988]. We modeled P&R for the VT (vertical transportation) task, based on the description made by Yost [Yost, 1994]. The VT task defines the design problem in which the goal is to configure an elevator. Figure 3 illustrates the inference structure for P&R abstracted from the VT domain. We decompose P&R into five primitive inferences.

  
Figure 3: Our representation of the inference structure for Propose&Revise. The inference structure specifies the inferences independent of their order of application.

• The select parameter inference chooses one parameter to have its value computed. The inference uses the INPUT PARAMETERS, the PARAMETER VALUES already computed and the PARAMETER DEPENDENCY RELATIONS to obtain a SELECTED PARAMETER (cf. section 4.1).
• The propose inference computes the value of the selected parameter. The inference uses the SELECTED PARAMETER and the PARAMETER PROCEDURES to compute new PARAMETER VALUES (cf. section 4.2).
• The check inference verifies the constraints after computing parameter values. The inference uses the new PARAMETER VALUES computed, the CONSTRAINT PROCEDURES, and the CONSTRAINT DEPENDENCY RELATIONS to compute the CONSTRAINT RESULTS (cf. section 4.3).
• The select violated constraint chooses one violated constraint to be revised. The inference uses the CONSTRAINT RESULTS and the FIX DEPENDENCY RELATIONS to generate one SELECTED VIOLATED CONSTRAINT (cf. section 4.4).
• The revise inference remedies a violated constraint. The inference uses the SELECTED VIOLATED CONSTRAINT, FIXES and FIX PROCEDURES to repair the constraint violation, and to propose new PARAMETER VALUES (cf. section 4.5).

The inference structure presents a general view of the problem-solving method and the knowledge used by the inferences. However, this description is abstract and imprecise. The information about the structure of knowledge roles is not revealed. For example, the parameters are related to one another by dependency relations, but there is no information about the organization of these dependency relations. In addition, we do not know if it is possible for a parameter to have more than one constraint, or if more than one fix for a constraint may exist.

In the next section, we propose an approach for describing formally the knowledge roles of a PSM in terms of a method ontology. This ontology unveils the assumptions about the structure of the knowledge roles required by a method. The advantage of describing a method ontology is a better understanding of a PSM, without going into details of the implementation.