for Building Expert Systems
Bing Leng, Bruce G. Buchanan · 1992
Induction programs make several assumptions that limit their practical utility. This paper reports on our research to overcome the limitation of working within a fixed vocabulary. A recently reported phenomenon in machine learning is that there is a tradeoff between the simplicity of concept descriptions and their coverage of trainging instances and a learning system cannot have both. We argue that if a learning system can generate new terms, it can achieve both simplicity and coverage. We give a new method for generating one kind of new terms, comparative terms. The experimental results on a mushroom classification task show that a single comparative term can achieve 90% predictive accuracy.