(An) efficient evolutionary optimization with fitness approximation using neural networks

Abstract

Evolutionary algorithms (EAs) usually need a large number of fitness function evaluations. In real-word applications, evaluating the fitness function is quite expensive and time consuming. To solve the above problems, fitness approximation is quite necessary for reducing the fitness evaluations. In this thesis, fitness function is approximated by incremental learning of a multilayer perceptron (MLP) network. A step based fitness approximation is proposed in this thesis. The whole fitness approximation is divided into three steps. The beginning certain generations and the normalized mean squared error (nMSE) of the approximate model are used to decide which step the fitness computation should be in. In step 1 and step 3, all of the individuals are evaluated with the original fitness function in each generation, and in step 2, certain individuals (estimated best and worst individuals) are revaluated by the original fitness function and the others are estimated by the approximate model. MLP network for fitness approximation is updated with the new available data. The proposed scheme is tested on five benchmark problems and compared with the fitness approximation without step based management. Simulation results show the proposed algorithm can reduce the original fitness function evaluations with the same optimization level.