Читать книгу Introduction to Fuzzy Logic - James K. Peckol - Страница 54

Learning by example or induction increases the level of participation by the learner in the learning process. Unlike the previous example in which the teacher abstracted and then presented the material, here the student must assume the responsibility for the task. In such a context, specific conceptual instances are presented, and the student must recognize the significant or key features of the examples and then form the desired generalizations.

An early classic example of such an approach is Patrick Winston's work on “Learning Structural Descriptions from Examples.” The goal of Winston's program was to learn elementary geometrical constructs such as those one might build using toy blocks. The program was presented with training instances from which it evolved an internal description of the concept it was to be learning. The knowledge acquired was incorporated into a semantic network where all of the interrelationships among the constituent elements were described.

Critical to the effective use of Winston's algorithm are the ideas that positive training instances are evolutionary rather than revolutionary. In Winston's algorithm, negative training instances are those that reflect only minimal differences from the concept being investigated; thus, no learning occurs.