This strategy represents a very simple but efficient way of producing the classification of an example given a set of potentially conflicting rules. It assumes that each rule is characterised by a value which expresses its "quality". When rules are generated by an inductive system this is easily obtained during the learning phase. Here we use a measure of quality provided by YAILS which is a function of two properties: its consistency and completeness. Rule quality is calculated as follows:
The notion of quality used here is a weighted sum of the consistency and completeness of the rule. The weights are proportional to the value of consistency giving thus some degree of flexibility (see [12, 13] for more details). Our formula for the calculation of quality is a heuristic one. Many other possibilities exist for evaluating a composite effect of various rule properties (see for instance [1] for a function which also includes simplicity).
Let us now come back to the best rule strategy. All rules applicable to a given example form a candidate set. After the candidate set has been formed, the rule with the highest quality value is chosen. The conclusion of this rule is followed.