Automatic selection of examples for training a learning design system

It has been observed that a careful selection of training examples improves the performance of supervised concept learning systems. This observation is important when examples are not available and are expensive to generate, or when using incremental learning systems. These two reasons are manifested in the context of BRIDGER, a system that learns to design cable-stayed bridges. This paper describes example selection techniques for the concept formation system ECOBWEB—the system that is used to acquire the synthesis knowledge in BRIDGER. The approach was implemented and tested. Preliminary results show that small improvements in syi uesis performance can be obtained by using the new technique.