A comparison of approaches to stepwise regression for global sensitivity analysis used with evolutionary optimization

Citation

Wang M, Wright JA, Buswell R & Brownlee A (2013) A comparison of approaches to stepwise regression for global sensitivity analysis used with evolutionary optimization. In: Proceedings of BS2013: 13th Conference of International Building Performance Simulation Association, Chambéry, France, August 26-28. BS2013: 13th Conference of International Building Performance Simulation Association, Chambéry, France, 26.08.2013-28.08.2013. London: International Building Performance Simulation Association, pp. 2551-2558. http://www.ibpsa.org/proceedings/BS2013/p_1047.pdf

Abstract
Applying global sensitivity analysis to solutions obtained from optimization gives an understanding of which variables have most impact on the solution. It also provides confidence in the optimality of the solution(s). Typically, global sensitivity analysis is based on a linear regression model in the stepwise manner. To improve computational efficiency, solutions obtained from optimization can be re-used to compute global sensitivities of variables. This paper investigates the extent to which the procedure options of stepwise regression analysis can influence the rank-order of variables importance, when using solutions taken from optimization. It is concluded that the stepwise regression analysis applied to rank transformed data from the first 100 optimization solutions, through bidirectional elimination and BIC, can rank the most important variables fast and accurately. In contrast, the production of more detailed information requires the use of AIC and larger sample sizes.

People

Dr Sandy Brownlee

Senior Lecturer in Computing Science, Computing Science and Mathematics - Division