Computational Intelligence in Optimization

Abstract

PrefaceOptimization is an integral part to science and engineering. Most real-world applications involve complex optimization processes, which are difficult to solve without advanced computational tools. With the increasing challenges of fulfilling optimization goals of current applications there is a strong drive to advance the development of efficient optimizers. The challenges introduced by emerging problems include:• objective functions which are prohibitively expensive to evaluate, so typically so only a small number of objective function evaluations can be made during the entire search, • objective functions which are highly multimodal or discontinuous, and • non-stationary problems which may change in time (dynamic).Classical optimizers may perform poorly or even may fail to produce any improvement over the starting vector in the face of such challenges. This has motivated researchers to explore the use computational intelligence (CI) to augment classical methods in tackling such challenging problems. Such methods include population-based search methods such as: a) evolutionary algorithms and particle swarm optimization and b) non-linear mapping and knowledge embedding approaches such as artificial neural networks and fuzzy logic, to name a few. Such approaches have been shown to perform well in challenging settings. Specifically, CI are powerful tools which offer several potential benefits such as: a) robustness (impose little or no requirements on the objective function) b) versatility (handle highly non-linear mappings) c) self-adaption to improve performance and d) operation in parallel (making it easy to decompose complex tasks). However, the successful application of CI methods to real-world problems is not straightforward and requires both expert knowledge and trial-and-error experiments. As such the goal of this volume is to survey a wide range of studies where CI has been successfully applied to challenging real-world optimization problems, while highlighting the VIII Preface insights researchers have obtained. Broadly, the studies in this volume focus on four main disciplines: continuous optimization, classification, scheduling and hardware implementations.For continuous optimization, Neto et al. study the use of artificial neural networks (ANNs) and Heuristic Rules for solving large scale optimization problems. They focus on a recurrent ANN to solve a quadratic programming problem and propose several techniques to accelerate convergence of the algorithm. Their method is more efficient than one using an ANN only. Starzyk et al. propose a direct-search optimization algorithm which uses reinforcement learning, resulting in an algorithm which 'learns' the best path during the search. The algorithm weights past steps based on their success to yield a new candidate search step. They benchmark their algorithm with several mathematical test functions and apply to training of a multi-layer perceptron neural network for image recognition. Ventresca et al. use the opposition sampling approach to decrease ...

Robust Turbine Blade Optimization in the Face of Real Geometric Variations

“…III. Moreover, as an improvement to related work on compressor blades, where deviations were approximated with a truncated multivariate normal distribution [8], probability density functions (PDFs) of random paa OPTIMAT v2 is a surrogate-based optimization tool developed at the University of Southampton's Rolls-Royce UTC [7]. rameters are estimated nonparametrically through kernel density estimation (KDE).…”