A comprehensive survey of fitness approximation in evolutionary computation

Abstract

Evolutionary algorithms (EAs) have received increasing interests both in the academy and industry. One main difficulty in applying EAs to real-world applications is that EAs usually need a large number of fitness evaluations before a satisfying result can be obtained. However, fitness evaluations are not always straightforward in many real-world applications. Either an explicit fitness function does not exist, or the evaluation of the fitness is computationally very expensive. In both cases, it is necessary to estimate the fitness function by constructing an approximate model. In this paper, a comprehensive survey of the research on fitness approximation in evolutionary computation is presented. Main issues like approximation levels, approximate model management schemes, model construction techniques are reviewed. To conclude, open questions and interesting issues in the field are discussed.Evolutionary computation has found a wide range of applications in various fields of science and engineering. Among others, evolutionary algorithms have been proved to be powerful global optimizers. Generally, evolutionary algorithms outperform conventional optimization algorithms for problems which are discontinuous, non-differential, multi-modal, noisy and not well-defined problems, such as art design, music composition and experimental designs [76]. Besides, evolutionary algorithms are also well suitable for multi-criteria problems.Despite the great successes achieved in real-world applications, evolutionary algorithms have also encountered many challenges. For most evolutionary algorithms, a large number of fitness evaluations (performance calculations) are needed before a well acceptable solution can be found. In many real-world applications, fitness evaluation is not trivial. There are several situations in which fitness evaluation becomes difficult and computationally efficient approximations of the fitness function have to be adopted.Several issues need to be addressed in employing fitness approximations in evolutionary computation. First, which levels of the fitness approximation should be used. While an experimental verification can be seen as the true fitness value of a given solution, fully computational simulations, simplified computational simulations as well as functional approximations (meta-models) can be used for fitness calculation. So far, several models have been used for fitness approximation. The most popular ones are polynomials (often known as response surface methodology), the kriging model, most popular in design and analysis of computer experiments (DACE), the feedforward neural networks, including multi-layer perceptrons and radialbasis-function networks and the support vector machines. Due to the lack of data and the high dimensionality of input space, it is very difficult to obtain a perfect global functional approximation of the original fitness function. To tackle this problem, two main measures can be taken.Firstly, the approximate model should be used together with the original fitness functi...

Comparison-Based Optimizers Need Comparison-Based Surrogates

“…Because Evolutionary Algorithms (EAs) require a lot of fitness evaluations, much work has been devoted in the last decade to specifically tune meta-model assisted approaches to EAs (see e.g. [9] for a survey -and section 2).…”

From evolutionary computation to the evolution of things

“…Increasingly EAs are being used for problems where evaluating each population member over many generations would take too long to permit effective evolution given the resources available. A range of approaches, collectively known as surrogate models, are being developed that use computationally cheaper models in place of full fitness evaluations, and refine those models via occasional full evaluations of targeted individuals 72,73,74,75 .…”

Improving Local Convergence in Particle Swarms by Fitness Approximation Using Regression

“…In recent years, several studies have proposed EAs incorporating fitness approximation with an aim to improve performance while not incurring expensive computational cost [15,23,28]. Typically these EAs employ a surrogate model in place of the expensive original function evaluations.…”

“…The most commonly used techniques for constructing surrogate models include Kriging [31], neural networks [17], and polynomial regression [23,32,37]. A recent survey on fitness approximation in EAs can be found in [15].…”