A self-adapting algorithm for many-objective optimization

Abstract

An increase in computing power available from modern processors and GPUs has allowed more complex design optimization problems to be solved. This increase in complexity typically results in a difficult optimization problem with a highly non-linear objective function topology. In this work a robust optimization algorithm is presented capable of efficiently traversing higher-dimensional objective function space to find Pareto optimal solutions. It is well known that no single optimization algorithm performs best for all problems. Therefore, the developed method, a many-objective hybrid optimizer (MOHO), uses five constitutive algorithms and actively switches among them throughout the optimization process to accelerate the convergence, avoid local minima and arrive at a diverse set of optimal designs. MOHO monitors the progress made by each of the five algorithms and allows the best performing algorithm more attempts at finding the optimum. The MOHO algorithm was tested on 13 unconstrained and five constrained analytical test problems, of up to 15 objectives, from the DTLZ and WFG test suite. The MOHO algorithm performed, on average, better than the five constitutive algorithms in 52% of the unconstrained DTLZ+WFG problems and in 70% of the constrained DTLZ test problems.