Solve least absolute value regression problems using modified goal programming techniques

Abstract

Least absolute value (LAV) regression methods have been widely applied in estimating regression equations. However, most of the current LAV methods are based on the original goal program developed over four decades. On the basis of a modified goal program, this study reformulates the LAV problem using a markedly lower number of deviational variables than used in the current LAV methods. Numerical results indicate that for the regression problems with hundreds of observations, this novel method can save more than 1/3 of the CPU time compared to current LAV methods.