Surrogate Modeling with Artificial Neural Networks for Efficient Optimization

Abstrakt

The optimization problems that are high dimensional occur frequently in complex engineering and scientific problems, Regardless of the size of the problem, the evaluation of the objective functions, particularly in complex simulations or black-box models is often computationally costly or impossible. Surrogate modeling provides a sensible alternative, approximating these black-box tasks which are expensive to model with data-driven approaches. This paper presents a surrogate modeling framework that can allow efficient optimization in high-dimensional design contexts through Artificial Neural Networks (ANNs). The DTLZ and ZDT benchmark function suites enabled one to stringently assess the model performance across a range of problem dimensions and complexity. We modeled feedforward neural networks to learn using initial data generating (Latin Hypercube Sampling, LHS) to capture the nonlinear interactions of design factors and objective functions. Our models demonstrated significant generalizing ability in validation datasets with a prediction rate of up to 98% (R 2 value) in 50 dimensional issue contexts. In addition to the ANN-based surrogates were the optimization methods, the optimization methods reduce the cost of the expensive function evaluation considerably in an effort to enhance the efficiency of computing without deteriorating the quality of the solution. These findings suggest that ANNs are viable strong surrogate models of high-dimensional optimization problems and that a scalable and accurate solution is possible to high-dimensional optimization problems in practice in scientific fields such as materials science, fluid dynamics, and aircraft design.