Learning by observation and active experimentation in a knowledge based CAD-environment

Abstract

A novel approach to the integration of machine learning into a knowledge-based CAD environment is presented. To achieve increased learning efficiency the learning system combines learning from observation during normal operation of the CAD system with active experimentation during its idle times. Automated example generation is based on metaknowledge about the design expertise implemented in the CAD system. By reusing specific parts of this knowledge to construct experiments the learning system automatically adapts to improvements and extensions in the host system. The current prototype was able to learn analytical knowledge about worst-case estimations for analog circuit blocks.<>