Recurrent neural network ensembles for convergence prediction in surrogate-assisted evolutionary optimization

Abstract

Evaluating the fitness of candidate solutions in evolutionary algorithms can be computationally expensive when the fitness is determined using an iterative numerical process. This paper illustrates how an ensemble of Recurrent Neural Networks can be used as a robust surrogate to predict converged Computational Fluid Dynamics data from unconverged data. The training of the individual neural networks is controlled and a variance range is used to determine if the surrogates have been adequately trained to predict diverse and accurate solutions. Heterogeneous ensemble members are used due to the limited data available and results show that for certain parameters, predictions can be made to within 5% of the converged data's final output, using approximately 40% of the iterations needed for convergence. The implications of the method and results presented are that it is possible to use ensembles of Recurrent Neural Networks to provide accurate fitness predictions for an evolutionary algorithm and that they could be used to reduce the time needed to achieve optimal designs based on time-consuming Computational Fluid Dynamics simulations.