Unsupervised learning of design rules aided by system derived heuristics

This paper describes HERALD (HEuristic Rule Analysis and Learning for Design), a two stage learning system consisting of ECMA (Extended Conceptual Modelling Assistant) and LEADER (LEArning DEsign Rules). HERALD provides design assistance for conceptual modelling in the Extended Entity-Relationship Model (ER) domain. Its first component, ECMA, using an extended case-based model and the expert as oracle, provides designers with suggestions for design completion by developing and refining heuristics reflecting modelling strategies. These heuristics are subsequently used by LEADER which, non-interactively, learns design rules via an explanation-based paradigm. Once approved by experts, LEADER'S rules are directly incorporated into the underlying knowledge-based design tool which HERALD supports. This work addresses what we perceive as a shortcoming of some AI-based design tools, which require detailed and explicit knowledge of design. In our approach the problem is circumvented by combining the explanation-based learning (EBL) with case-based learning. The latter type of learning, rather than requiring a codified theory of design, uses readily available design cases.