SPARSE VARIABLE SELECTION VIA ITERATIVELY REWEIGHTED LEAST SQUARES

Recently, there has been growing interest in computer experiments which are extremely important in modern industry, science, and engineering. A feature of computer experiments is that the number of input variables can be quite large, so screening active-effective factors is very important. One of the goals in computer experiments is to choose a parsimonious model by variable selection that is viewed as a kind of regularization. In this paper, a novel algorithm is proposed originated from signal processing. Different from the existed methods, we present a reweighted L-2 norm to approximate the L-p(0 < p < 1) minimization problems with good sparseness for variable selection. The algorithm consists of solving a sequence of weighted L-2 minimization problems where the weights used for the next iteration are computed from the value of the current solution. This novel regularization strategy which has no necessity to choose the proper regularization parameter is found to greatly improve the ability of variable selection. We illustrate our algorithm by a famous example and make comparison with previous works. Experimental results demonstrate that the new algorithm not only behaves better than many other methods, bust also has low computational cost.