constructive induction in engineering design

(Arciszewski, Michalski, Wnek, Bloedorn)

The ultimate objective of this project is to develop a class of constructive induction methods for the applications to engineering design and a practical methodology for their use. A feasibility study has been completed and its results presented in the research report (Arciszewski et al 1992) published at the Center for Artificial Intelligence at George Mason University and in the ASCE Journal of Computing in Civil Engineering (Arciszewski et al, P94-16). The study was conducted in the area of conceptual design of wind bracings in steel skeleton structures of tall buildings.

Design rules were learned from a collection of 336 examples of minimum weight (optimal) designs of wind bracings. Constructive induction was used to produce design rules which explain how design requirements can be optimally (in terms of minimum steel weight) satisfied through the proper selection of individual components of a wind bracing structural system. All examples were prepared under identical design assumptions for a three-bay skeleton of a tall building in cooperation with practicing structural designers. Actual minimum-weight designs were produced using SODA, a computer system for optimization, analysis, and design of steel structures. The design rules obtained were divided into four classes corresponding to the value of the decision attribute: recommendation, standard, avoidance and infeasibility rules.

Two types of constructive induction have been used in the study: data-driven and hypothesis-driven constructive induction. The performance of both learning systems was formally measured by two empirical error rates: 1. the overall empirical error rate, 2. the omission error rate in accordance to the method of evaluation of performance of learning systems developed at the Laboratory and published in Arciszewski et al (1994). These error rates were calculated for the entire collection of examples using the leave-one-out resampling method. The error rates for constructive induction were compared with rates for the “traditional” induction, based on the use of the AQ15 algorithm. The individual error rates are shown in the table below. There is a significant improvement in performance (more than 50%) between the system based on the “traditional” induction and systems based on constructive induction. The difference in performance between two constructive induction-based systems is insignificant (less than 5%), but this may change as the research progresses.

Selected References

Chen, Q. and Arciszewski, T., “Machine Learning of Bridge Design Rules: A Case Study,” Proceedings of the 2nd ASCE Congress on Computing in Civil Engineering, Atlanta GA, June, 1995.

Arciszewski, T., Bloedorn, E., Michalski, R.S., Mustafa, M. and Wnek, J., “Machine Learning of Design Rules: Methodology and Case Study,” ASCE Journal of Computing in Civil Engineering, Vol. 8, No. 3, pp. 286-308, July 1994.

Arciszewski, T., “Machine Learning in Engineering Design,” Proceedings of the Conference on Intelligent Information Systems, Institute of Computer Science, Polish Academy of Sciences, Wigry, Poland, 1994.

Wnek, J., Michalski, R.S. and Arciszewski, T., “An Application of Constructive Induction to Engineering Design,” Proceedings of the IJCAI-93 Workshop on AI in Design, Chambery France, August 1993.

Arciszewski, T., Ziarko, W. and Khan, T.L., “Learning Conceptual Design Rules: A Rough Sets Approach,” Proceedings of the International Workshop on Rough Sets, Banff, Alberta, Canada, 1993.

Arciszewski, T., Bloedorn, E., Michalski, R.S., Mustafa, M. and Wnek, J., “Constructive Induction in Structural Design,” Reports of the Machine Learning and Inference Laboratory, School of Information Technology and Engineering, George Mason University, MLI 92-07, December 1992.

For more references, seeĀ publications section.