Abstract: Multidimensional scaling (MDS) algorithms are widely used in node localization for wireless sensor networks by constructing a pair-wise squared distance matrix and performing a double-centered transformation. Classical MDS algorithms reconfigure relative coordinates of nodes in a similarity space and give the solution based on least squares (LS) criterion. However, the transformation of classical MDS algorithms result in non-Gaussian distribution of the noise in the similarity matrix when white Gaussian noise exists in distance measurements. Thus the LS based estimator can not optimize the node location. To overcome this problem, a least absolute deviation (LAD) based cost function of classical MDS algorithm is presented. Simulation results show that the LAD based method yields better performance.