Optimization of simulation model performance requires finding the values of the model's controllable inputs that optimize a chosen model response. Responses are usually stochastic in nature, and the cost of simulation model runs is high. The literature suggests the use of metamodels to synthesize the response surface using sample data. In particular, nonparametric regression is proposed as a useful tool in the global optimization of a response surface. As the general simulation optimization problem is very difficult and requires expertise from a number of fields, there is a growing consensus in the literature that a knowledge-based approach to solving simulation optimization problems is required. This dissertation examines the relative performance of the principal nonparametric techniques, spline and kernel smoothing, and subsequently addresses the issues involved in implementing the techniques in a knowledge-based simulation optimization system. The dissertation consists of two parts. In the first part, a full factorial experiment is carried out to compare the performance of kernel and spline smoothing on a number of measures when modeling a varied set of surfaces using a range of small sample sizes. In the second part, nonparametric metamodeling techniques are placed in a taxonomy of stochastic search procedures for simulation optimization and a method for their implementation in a knowledge-based system is presented. A sequential design procedure is developed that allows spline smoothing to be used as a search technique. Throughout the dissertation, a two-input, single-response model is considered. Results from the experiment show that spline smoothing is superior to constant-bandwidth kernel smoothing in fitting the response. Kernel smoothing is shown to be more accurate in placing optima in X-space for sample sizes up to 36. Inventory model examples are used to illustrate the results. The taxonomy implies that search procedures can be chosen initially using the parameters of the problem. A process that allows for selection of a search technique and its subsequent evaluation for further use or for substitution of another search technique is given. The success of a sequential design method for spline smooths in finding a global optimum is demonstrated using a bimodal response surface. / Ph. D.

When creating a muffler to be installed on a noise generating machine, the design parameters as well as the placements of sound attenuating elements has to be optimized in order to minimize the sound coming out of the equipage. This is exemplified in a small project task for students of a basic course in optimization at KTH. The task is however flawed, since both the way in which the optimization problem is formed is overly simplistic and the algorithm used to solve the problem, fmincon, does not cope well with the mathematical complexity of the model, meaning it gets stuck in a local optimum that is not a global optimum. This thesis is about investigating how to solve both of these problems. The model is modified to combine several frequencies and adjusting them to the sensitivity to different frequencies in the human ear. By doing this, the objective is changed from the previous way of maximizing Dynamic Insertion Loss Dilfor a specific frequency to minimize the total perceived sound level LA.  The model is based on the modular design of TMM from four-pole theory. This divides the muffler into separate parts, with the sound attenuating elements being mathematically defined only by what T matrix it has. The element types to choose from are the Expansion Chamber, the Quarter Wave Resonator and the Helmholtz Resonator. The global optimization methods to choose from are Global Search, MultiStart, Genetic Algorithm, Pattern Search and Simulated Annealing. By combining the different types of sound attenuating elements in every way and solving each case with every global optimization method, the best combination to implement to the model is chosen. The choice is two Quarter Wave Resonators being solved by MultiStart, which provides satisfactory results. Further analysis is done to ensure the robustness of chosen implementation, which does not reveal any significant flaws. The purpose of this thesis is fulfilled. / När man skapar en ljuddämpare som ska installeras på en ljud-genererande maskin bör designparametrarna samt placeringarna av ljuddämpande element optimeras för att minimera ljudet som kommer ut ur ekipaget. Detta exemplifieras i en liten projektuppgift för studenter till en grundkurs i optimering på KTH. Uppgiften är dock bristfällig, eftersom både det sätt som optimeringsproblemet är utformat är alltför förenklat och den algoritm som används för att lösa problemet, fmincon, inte klarar av modellens matematiska komplexitet bra, vilket menas med att den fastnar i ett lokalt optimum som inte är ett globalt optimum. Detta examensarbete handlar om att undersöka hur man kan lösa båda dessa problem. Modellen är modifierad för att kombinera flera frekvenser och anpassa dem till känsligheten för olika frekvenser i det mänskliga örat. Genom att göra detta är målet ändrat från det tidigare sättet att maximera den dynamiska insatsisoleringen DIL för en specifik frekvens till att minimera den totala upplevda ljudnivån LA. Modellen bygger på den modulära designen av TMM från 4-polsteori. Detta delar upp ljuddämparen i separata delar, med ljuddämpande element som matematiskt endast definieras av vilken T matris de har. De elementtyper att välja mellan är expansionskammare, kvartsvågsresonator och Helmholtzresonator. De globala optimeringsmetoder att välja mellan är Global Search, MultiStart, Genetic Algorithm, Pattern Search och Simulated Annealing. Genom att kombinera de olika typerna av ljuddämpande element på alla sätt och lösa varje fall med varje global optimeringsmetod, blir den bästa kombinationen vald och implementerad i modellen. Valet är två kvartsvågsresonatorer som löses genom MultiStart, vilket ger tillfredsställande resultat. Ytterligare analyser görs för att säkerställa robustheten av den valda implementationen, som inte avslöjar några väsentliga brister. Syftet med detta examensarbete är uppfyllt.

Today the modern power system is more dynamic and its operation is a subject to a number of constraints that are reflected in various management and planning tools used by system operators. In the case of hourly generation planning, Economic Dispatch (ED) allocates the outputs of all committed generating units, which are previously identified by the solution of the Unit Commitment (UC) problem. Thus, the accurate solutions of the ED and UC problems are essential in order to operate the power system in an economic and efficient manner. A number of computation techniques have progressively been proposed to solve these critical issues. One of them is a Particle Swarm Optimisation (PSO), which belongs to the evolutionary computation techniques, and it has attracted a great attention of the research community since it has been found to be extremely effective in solving a wide range of engineering problems. The attractive characteristics of PSO include: ease of implementation, fast convergence compared with the traditional evolutionary computation techniques and stable convergence characteristic. Although the PSO algorithms can converge very quickly towards the optimal solutions for many optimisation problems, it has been observed that in problems with a large number of suboptimal areas (i.e. multi-modal problems), PSO could get trapped in those local minima, including ED and UC problems. Aiming at enhancing the diversity of the traditional PSO algorithms, this thesis proposes a method of combining the PSO algorithms with a real-valued natural mutation (RVM) operator to enhance the global search capability and investigate the performance of the proposed algorithm compared with the standard PSO algorithms and other algorithms. Prior to applying to ED and UC problems, the proposed method is tested with some selected mathematical functions where the results show that it can avoid being trapped in local minima. The proposed methodology is then applied to ED and UC problems, and the obtained results show that it can provide solutions with good accuracy and stable convergence characteristic with simple implementation and satisfactory calculation time. Furthermore, the sensitivity analysis of PSO parameters has been studied so as to investigate the response of the proposed method to the parameter variations, especially in both ED and UC problems. The outcome of this research shows that the proposed method succeeds in dealing with the PSO' s drawbacks and also shows the superiority over the traditional PSO algorithms and other methods in terms of high quality solutions, stable convergence characteristic, and robustness.

Les (sub)nanoclusters sont des agrégats d’atomes ou de molécules composés de quelques unités à quelques centaines d’unités. En raison de leur petite taille, ils peuvent avoir des propriétés électroniques, optiques, magnétiques et catalytiques très différentes par rapport au solide correspondant . D'un point de vue expérimental, il est encore très difficile de synthétiser des agrégats de taille calibrée. D'un point de vue théorique, le développement des puissances de calcul, des méthodes de calcul de structure électronique et des algorithmes de recherches globales de structures stables, permettent un calcul toujours plus précis de leurs propriétés physico-chimiques. L’étude théorique permet alors de déterminer de façon fiable les structures stables de ces systèmes qui président aux calculs de leurs propriétés . L’exemple qui illustre ce travail s’inspire du processus observé au sein des piles à combustible dans lequel le Platine (Pt) est couramment utilisé pour produire de l’énergie par oxydation du dihydrogène en favorisant notamment sa dissociation . L’objet de ce travail consiste à comparer la capacité des clusters de Platine de différentes tailles à adsorber la molécule de dihydrogène sous leur forme libre et adsorbée sur substrat. Le graphène , matériaux bidimensionnel cristallin formé de carbone est choisi dans ce travail en tant que substrat en raison de sa grande résistance mécanique et chimique. La première partie de ce travail est consacrée à la recherche d’éléments dopants qui vont permettent à la fois d’améliorer la capacité d’adsorption des clusters de Platine sur la surface et éviter leur migration. L’objectif est ici de proposer un substrat sur lequel peuvent être empêchés les phénomènes d’agglomération, de dissolution et de détachement du cluster qui ainsi limiteraient son efficacité catalytique . Des dopages de la surface, tel qu’ils sont réalisables expérimentalement , par l’Azote, le Bore et le Nitrure de Bore, par substitution atomique et avec ou sans considération préalable de lacunes, ont été étudiés. La seconde partie correspond à l’implémentation dans le code GSAM (Global Search Algorithm of Minima - algorithme de recherche globale de minima) développé au laboratoire , , des éléments qui permettent la recherche de structures de plus basse énergie de clusters moléculaires adsorbés sur substrat, tels que les systèmes [H2-Ptn-Graphène dopé] de cet exemple. La troisième partie concerne l’illustration de la fiabilité de la méthode de recherche globale employée et de la qualité de quelques méthodes de calcul de l’énergie moléculaire (DFT et GUPTA) vis-à-vis de résultats mentionnés dans la littérature sur les clusters de Platine. La dernière partie comporte l’investigation structurale des systèmes [H2-Ptn] et [H2-Ptn-Graphène dopé] pour différentes tailles de clusters allant de n=6 à n=20. La variation de l’énergie d’adsorption de H2 sur les clusters libres et supportés ainsi que celle du cluster moléculaire sur le substrat en fonction de la taille est reportée. / A sub-nanometer sized metal cluster consists of only several to tens of atoms. Due to its small size and quantum effects, it can have very specific electronic, optical, magnetic and catalytic properties as compared with their bulk behaviors . From an experimental point of view, it is still a big challenge to realize size-controlled synthesis for (sub) nanoclusters. From a theoretical point of view, benefiting from the development of faster high-performance computational sources, more efficient electronic structure modelling software and more reliable global search methods for the determination of the most stable structures, the chemical and physical properties of clusters can be determinate more accurately. As it is experimentally a big challenge to realize size-controlled synthesis for (sub) nanoclusters, theoretical studies can provide detailed information on their geometric structure, electronic structure, as well as adsorption and reaction properties . The example chosen to be treated in this study is inspired by the fuel cell, in which the Platinum (Pt) is a typical and most commonly used precious metal catalyst for the production of energy by the oxidation of dihydrogene . Graphene is a recently discovered 2D carbon net structure, has several special properties, such as: low weight, high strength, high surface area, high electrical conductivity, etc. With these properties and their novel combinations, graphene might prove a promising candidate to be used as catalyst supports. The first part of this study is devoted to the search of the doping elements which permit both enhance the adsorption capacity of Pt clusters on the surface and prevent their migration. The aim here is propose one substrate which can avoid the problems of cluster agglomeration, dissolution and detachment, which reduce the performance of the catalysts . The ways of doping of the surface, which have already been experimentally realized , such as Nitrogen, Boron, and N-B patches substitution of Carbon atoms with or without introducing the vacancy on the pristine graphene, are studied. The second part corresponds to the implementation of some new features into the code GSAM (Global Search Algorithm of Minima) developed in our laboratory , , , which permit the search of the most stable structures of the molecular clusters adsorbed on substrate, such as the complex systems of [H2-Ptn-doped Graphene]. The third part is to evaluate the reliabilities of the global search method used, as well as the DFT and the empirical (GUPTA) potential energy surface. Thus, the main discussion appears as a comparison with the results of the literature concerning the Pt clusters. The fourth part consists of the structural investigation of [H2-Ptn] and [H2-Ptn-doped Graphene] systems for different sizes of Pt clusters with n=6 to n=20. The variation of the adsorption energy of H2 on the free and supported Ptn clusters, and the adsorption energy of (H2+Ptn) system on the surface with respect to the size of the cluster is discussed.

In this paper we view learning as an unconstrained non-linear minimization problem in which the objective function is defined by the negative log-likelihood function and the search space by the parameter space of an origin constrained product unit neural spatial interaction model. We consider Alopex based global search, as opposed to local search based upon backpropagation of gradient descents, each in combination with the bootstrapping pairs approach to solve the maximum likelihood learning problem. Interregional telecommunication traffic flow data from Austria are used as test bed for comparing the performance of the two learning procedures. The study illustrates the superiority of Alopex based global search, measured in terms of Kullback and Leibler's information criterion.

Local search algorithms have been proved to be effective in refining solutions that have been found by other algorithms. Evolutionary algorithms, in particular global search algorithms, have shown to be successful in producing approximate solutions for optimisation and classification problems in acceptable computation times. A relatively new method, memetic algorithms, uses local search to refine the approximate solutions produced by global search algorithms. This thesis develops such a memetic algorithm. The global search algorithm used as part of the new memetic algorithm is a genetic program that implements the building block hypothesis by building simplistic decision trees representing valid solutions, and gradually increases the complexity of the trees. The specific building block hypothesis implementation is known as the building block approach to genetic programming, BGP. The effectiveness and efficiency of the new memetic algorithm, which combines the BGP algorithm with a local search algorithm, is demonstrated. / Dissertation (MSc)--University of Pretoria, 2006. / Computer Science / unrestricted