An improved stochastic local search method in a multistart framework

UNIRANDI is a stochastic type, direct search method, where in each step a line search is started from the current point along a good random direction. In this paper we consider an improved version of this local search algorithm. The new method alters the random directions with new directions that rely on information from the previous stages of the optimization process. The improved algorithm was tested in a simple multistart framework and also as part of the GLOBAL method. The new local search technique is empirically compared with the old one on a unimodal and a multimodal function testbed up to moderate size dimensions.