A multiple surrogate assisted multi/many-objective multi-fidelity evolutionary algorithm

Engineering design commonly involves optimization of multiple conflicting performance objectives. During the optimization process, the performance of each candidate design/solution is evaluated using a model which may be empirical, numerical, experimental, etc., among other forms. The accuracy of the underlying model in representing the real-world behavior is referred to as fidelity. A low-fidelity model may be quick to evaluate but not very accurate; whereas a high-fidelity model may be computationally expensive to evaluate but provides an accurate estimate of the true performance. The paradigm of utilizing the low and high-fidelity models' information to identify the high-fidelity optimal solution(s) is known as multi fidelity optimization. This study delves into multi-fidelity optimization for problems which contain multiple objectives and where iterative solvers such as finite element analysis, computational fluid dynamics, etc. are used for performance evaluation. By stopping the solver at various stages before convergence, lower-fidelity performance estimates can be obtained at reduced computational cost. Most of the existing multi-fidelity methods can only deal with two fidelities (high and low) and a single objective. To overcome this research gap, we present a novel multi-objective evolutionary algorithm that can deal with multiple (arbitrary) number of fidelities by effectively utilizing pre-converged low-fidelity information. The proposed algorithm uses multiple surrogate models to capture the underlying function(s) with enhanced precision. A decomposition-based scheme is deployed for improved scalability in higher number of objectives. A classifier assisted pre-selection method is used to screen potential non-dominated solutions for efficient use of the computational budget. Additionally, a set of multi-fidelity, multi/many objective benchmark problems with different Pareto front types is also introduced to aid a systematic benchmarking. Numerical experiments are presented to highlight the efficacy of the proposed approach. (C) 2019 Elsevier Inc. All rights reserved.