A new algorithm for the problem of robust single objective optimization

Abstract

This paper propounds a new algorithm, the Sub-Space Random Search (SSRS) for the problem of singleobjective optimization, with the aim of improving the robustness and the precision of classical methods of global optimization. The new algorithm is compared with a genetic algorithm (GA), on a set of four scaleable test functions and with the number of variables changing from 1 to 5. A new test function called Deceptivebimodal (DB) is proposed. Results indicate that, with the same total number of function evaluations, SSRS is about 50% faster than GA. Moreover, SSRS shows a greater precision and similar ability to find the global optimum than GA with 1, 2 and sometimes 3 variables. But this advantage diminishes when the number of variables increases on multimodal and narrow-flat valley functions. Finally, SSRS is successfully applied to a problem of dynamical synthesis of a mechanism.