Learning Conceptual Design Rules: A Rough Sets Approach

The paper presents the results of a feasibility study conducted in the area of learning conceptual design rules governing the selection of wind brac­ ing components in steel skeleton structures of tall buildings. The study's objectives were to compare decision rules produced by different learning systems using the same body of examples, and to formally verify these rules using the overall empirical error rate. The study was conducted using two learning systems, both based on the theory of rough sets: 1) System ROUGH which usually produces a Jarge number of complex de­ terministic rules, 2) System DataLogic which can generate probabilistic rules, relatively simple and much fewer in number than in the case of ROUGH. All experiments were conducted using a coUection of 314 ex­ a.mples of minimum weight (optimal) design of wind br&cings in steel skeleton structures of tall buildings. The examples were prepared under identical design a.88umptions for a three bay skeleton structure of a tall building. They were produced using SODA, a. computer software package for the analysis, design and optimization of steel structures. The paper gives a description of the learning experiments performed. It also pro­ vides a comparison of decision rules produced by DataLogic and Rough. and an analysis of empirical error rates obtained for the various coUection of exa.mples for ROUGH.