<p>Practical design problems in engineering and science involve the evaluation of expensive black-box functions, the optimization of multiple—often conflicting—targets, and the integration of data generated by multiple sources of information, e.g., numerical models with different levels of fidelity. If not properly handled, the complexity of these design problems can lead to lengthy and costly development cycles. In the last years, Bayesian optimization has emerged as a powerful alternative to solve optimization problems that involve the evaluation of expensive black-box functions. Bayesian optimization has two main components: a probabilistic surrogate model of the black-box function and an acquisition function that drives the optimization. Its ability to find high-performance designs within a limited number of function evaluations has attracted the attention of many fields including the engineering design community. The practical relevance of strategies with the ability to fuse information emerging from different sources and the need to optimize multiple targets has motivated the development of multi-fidelity modeling techniques and multi-objective Bayesian optimization methods. A key component in the vast majority of these methods is the Gaussian process (GP) due to its flexibility and mathematical properties.</p> <p><br></p> <p>The objective of this dissertation is to develop new approaches in the areas of multi-fidelity modeling and multi-objective Bayesian optimization. To achieve this goal, this study explores the use of linear and non-linear compositions of GPs to build probabilistic models for Bayesian optimization. Additionally, motivated by the rationale behind well-established multi-objective methods, this study presents a novel acquisition function to solve multi-objective optimization problems in a Bayesian framework. This dissertation presents four contributions. First, the auto-regressive model, one of the most prominent multi-fidelity models in engineering design, is extended to include informative mean functions that capture prior knowledge about the global trend of the sources. This additional information enhances the predictive capabilities of the surrogate. Second, the non-linear auto-regressive Gaussian process (NARGP) model, a non-linear multi-fidelity model, is integrated into a multi-objective Bayesian optimization framework. The NARGP model offers the possibility to leverage sources that present non-linear cross-correlations to enhance the performance of the optimization process. Third, GP classifiers, which employ non-linear compositions of GPs, and conditional probabilities are combined to solve multi-objective problems. Finally, a new multi-objective acquisition function is presented. This function employs two terms: a distance-based metric—the expected Pareto distance change—that captures the optimality of a given design, and a diversity index that prevents the evaluation of non-informative designs. The proposed acquisition function generates informative landscapes that produce Pareto front approximations that are both broad and diverse.</p>

Several exact and approximate dynamic programming formulations have already been proposed to solve hotel revenue management (RM) problems. To obtain tractable solutions, these methods are often bound by simplifying assumptions which prevent their application on large and dynamic complex systems. This dissertation introduces HotelSimu, a flexible simulation-based optimization approach for hotel RM, and investigates possible approaches to increase the efficiency of black-box optimization methods in the presence of noise. In fact, HotelSimu employs black-box optimization and stochastic simulation to find the dynamic pricing policy which is expected to maximize the revenue of a given hotel in a certain period of time. However, the simulation output is noisy and different solutions should be compared in a statistically significant manner. Various black-box heuristics based on variations of random local search are investigated and integrated with statistical analysis techniques in order to manage efficiently the optimization budget.

Les Algorithmes Évolutionnaires (AEs) ont été très étudiés en raison de leur capacité à résoudre des problèmes d'optimisation complexes en utilisant des opérateurs de variation adaptés à des problèmes spécifiques. Une recherche dirigée par une population de solutions offre une bonne robustesse par rapport à un bruit modéré et la multi-modalité de la fonction optimisée, contrairement à d'autres méthodes d'optimisation classiques telles que les méthodes de quasi-Newton. La principale limitation de AEs, le grand nombre d'évaluations de la fonction objectif,pénalise toutefois l'usage des AEs pour l'optimisation de fonctions chères en temps calcul.La présente thèse se concentre sur un algorithme évolutionnaire, Covariance Matrix Adaptation Evolution Strategy (CMA-ES), connu comme un algorithme puissant pour l'optimisation continue boîte noire. Nous présentons l'état de l'art des algorithmes, dérivés de CMA-ES, pour résoudre les problèmes d'optimisation mono- et multi-objectifs dans le scénario boîte noire.Une première contribution, visant l'optimisation de fonctions coûteuses, concerne l'approximation scalaire de la fonction objectif. Le meta-modèle appris respecte l'ordre des solutions (induit par la valeur de la fonction objectif pour ces solutions); il est ainsi invariant par transformation monotone de la fonction objectif. L'algorithme ainsi défini, saACM-ES, intègre étroitement l'optimisation réalisée par CMA-ES et l'apprentissage statistique de meta-modèles adaptatifs; en particulier les meta-modèles reposent sur la matrice de covariance adaptée par CMA-ES. saACM-ES préserve ainsi les deux propriété clé d'invariance de CMA-ES: invariance i) par rapport aux transformations monotones de la fonction objectif; et ii) par rapport aux transformations orthogonales de l'espace de recherche.L'approche est étendue au cadre de l'optimisation multi-objectifs, en proposant deux types de meta-modèles (scalaires). La première repose sur la caractérisation du front de Pareto courant (utilisant une variante mixte de One Class Support Vector Machone (SVM) pour les points dominés et de Regression SVM pour les points non-dominés). La seconde repose sur l'apprentissage d'ordre des solutions (rang de Pareto) des solutions. Ces deux approches sont intégrées à CMA-ES pour l'optimisation multi-objectif (MO-CMA-ES) et nous discutons quelques aspects de l'exploitation de meta-modèles dans le contexte de l'optimisation multi-objectif.Une seconde contribution concerne la conception d'algorithmes nouveaux pour l'optimi\-sation mono-objectif, multi-objectifs et multi-modale, développés pour comprendre, explorer et élargir les frontières du domaine des algorithmes évolutionnaires et CMA-ES en particulier. Spécifiquement, l'adaptation du système de coordonnées proposée par CMA-ES est coupléeà une méthode adaptative de descente coordonnée par coordonnée. Une stratégie adaptative de redémarrage de CMA-ES est proposée pour l'optimisation multi-modale. Enfin, des stratégies de sélection adaptées aux cas de l'optimisation multi-objectifs et remédiant aux difficultés rencontrées par MO-CMA-ES sont proposées. / Evolutionary Algorithms (EAs) have received a lot of attention regarding their potential to solve complex optimization problems using problem-specific variation operators. A search directed by a population of candidate solutions is quite robust with respect to a moderate noise and multi-modality of the optimized function, in contrast to some classical optimization methods such as quasi-Newton methods. The main limitation of EAs, the large number of function evaluations required, prevents from using EAs on computationally expensive problems, where one evaluation takes much longer than 1 second.The present thesis focuses on an evolutionary algorithm, Covariance Matrix Adaptation Evolution Strategy (CMA-ES), which has become a standard powerful tool for continuous black-box optimization. We present several state-of-the-art algorithms, derived from CMA-ES, for solving single- and multi-objective black-box optimization problems.First, in order to deal with expensive optimization, we propose to use comparison-based surrogate (approximation) models of the optimized function, which do not exploit function values of candidate solutions, but only their quality-based ranking.The resulting self-adaptive surrogate-assisted CMA-ES represents a tight coupling of statistical machine learning and CMA-ES, where a surrogate model is build, taking advantage of the function topology given by the covariance matrix adapted by CMA-ES. This allows to preserve two key invariance properties of CMA-ES: invariance with respect to i). monotonous transformation of the function, and ii). orthogonal transformation of the search space. For multi-objective optimization we propose two mono-surrogate approaches: i). a mixed variant of One Class Support Vector Machine (SVM) for dominated points and Regression SVM for non-dominated points; ii). Ranking SVM for preference learning of candidate solutions in the multi-objective space. We further integrate these two approaches into multi-objective CMA-ES (MO-CMA-ES) and discuss aspects of surrogate-model exploitation.Second, we introduce and discuss various algorithms, developed to understand, explore and expand frontiers of the Evolutionary Computation domain, and CMA-ES in particular. We introduce linear time Adaptive Coordinate Descent method for non-linear optimization, which inherits a CMA-like procedure of adaptation of an appropriate coordinate system without losing the initial simplicity of Coordinate Descent.For multi-modal optimization we propose to adaptively select the most suitable regime of restarts of CMA-ES and introduce corresponding alternative restart strategies.For multi-objective optimization we analyze case studies, where original parent selection procedures of MO-CMA-ES are inefficient, and introduce reward-based parent selection strategies, focused on a comparative success of generated solutions.

The thesis compares several non-linear regression methods on synthetic data sets gen- erated using standard benchmarks for a continuous black-box optimization. For that com- parison, we have chosen the following regression methods: radial basis function networks, Gaussian processes, support vector regression and random forests. We have also included polynomial regression which we use to explain the basic principles of regression. The com- parison of these methods is discussed in the context of black-box optimization problems where the selected methods can be applied as surrogate models. The methods are evalu- ated based on their mean-squared error and on the Kendall's rank correlation coefficient between the ordering of function values according to the model and according to the function used to generate the data. 1

Surrogate-model-based optimization is widely used to solve black-box optimization problems if the evaluation of a target system is expensive. However, when the optimization budget is limited to a single or several evaluations, surrogate-model-based optimization may not perform well due to the lack of knowledge about the search space. In this case, transfer learning helps to get a good optimization result due to the usage of experience from the previous optimization runs. And if the budget is not strictly limited, transfer learning is capable of improving the final results of black-box optimization. The recent work in surrogate-model-based optimization showed that using multiple surrogates (i.e., applying multi-surrogate-model optimization) can be extremely efficient in complex search spaces. The main assumption of this thesis suggests that transfer learning can further improve the quality of multi-surrogate-model optimization. However, to the best of our knowledge, there exist no approaches to transfer learning in the multi-surrogate-model context yet. In this thesis, we propose an approach to transfer learning for multi-surrogate-model optimization. It encompasses an improved method of defining the expediency of knowledge transfer, adapted multi-surrogate-model recommendation, multi-task learning parameter tuning, and few-shot learning techniques. We evaluated the proposed approach with a set of algorithm selection and parameter setting problems, comprising mathematical functions optimization and the traveling salesman problem, as well as random forest hyperparameter tuning over OpenML datasets. The evaluation shows that the proposed approach helps to improve the quality delivered by multi-surrogate-model optimization and ensures getting good optimization results even under a strictly limited budget.:1 Introduction 1.1 Motivation 1.2 Research objective 1.3 Solution overview 1.4 Thesis structure 2 Background 2.1 Optimization problems 2.2 From single- to multi-surrogate-model optimization 2.2.1 Classical surrogate-model-based optimization 2.2.2 The purpose of multi-surrogate-model optimization 2.2.3 BRISE 2.5.0: Multi-surrogate-model-based software product line for parameter tuning 2.3 Transfer learning 2.3.1 Definition and purpose of transfer learning 2.4 Summary of the Background 3 Related work 3.1 Questions to transfer learning 3.2 When to transfer: Existing approaches to determining the expediency of knowledge transfer 3.2.1 Meta-features-based approaches 3.2.2 Surrogate-model-based similarity 3.2.3 Relative landmarks-based approaches 3.2.4 Sampling landmarks-based approaches 3.2.5 Similarity threshold problem 3.3 What to transfer: Existing approaches to knowledge transfer 3.3.1 Ensemble learning 3.3.2 Search space pruning 3.3.3 Multi-task learning 3.3.4 Surrogate model recommendation 3.3.5 Few-shot learning 3.3.6 Other approaches to transferring knowledge 3.4 How to transfer (discussion): Peculiarities and required design decisions for the TL implementation in multi-surrogate-model setup 3.4.1 Peculiarities of model recommendation in multi-surrogate-model setup 3.4.2 Required design decisions in multi-task learning 3.4.3 Few-shot learning problem 3.5 Summary of the related work analysis 4 Transfer learning for multi-surrogate-model optimization 4.1 Expediency of knowledge transfer 4.1.1 Experiments’ similarity definition as a variability point 4.1.2 Clustering to filter the most suitable experiments 4.2 Dynamic model recommendation in multi-surrogate-model setup 4.2.1 Variable recommendation granularity 4.2.2 Model recommendation by time and performance criteria 4.3 Multi-task learning 4.4 Implementation of the proposed concept 4.5 Conclusion of the proposed concept 5 Evaluation 5.1 Benchmark suite 5.1.1 APSP for the meta-heuristics 5.1.2 Hyperparameter optimization of the Random Forest algorithm 5.2 Environment setup 5.3 Evaluation plan 5.4 Baseline evaluation 5.5 Meta-tuning for a multi-task learning approach 5.5.1 Revealing the dependencies between the parameters of multi-task learning and its performance 5.5.2 Multi-task learning performance with the best found parameters 5.6 Expediency determination approach 5.6.1 Expediency determination as a variability point 5.6.2 Flexible number of the most similar experiments with the help of clustering 5.6.3 Influence of the number of initial samples on the quality of expediency determination 5.7 Multi-surrogate-model recommendation 5.8 Few-shot learning 5.8.1 Transfer of the built surrogate models’ combination 5.8.2 Transfer of the best configuration 5.8.3 Transfer from different experiment instances 5.9 Summary of the evaluation results 6 Conclusion and Future work

On s'intéresse à l'étude d'algorithmes stochastiques pour l'optimisation numérique boîte-noire. Dans la première partie de cette thèse, on présente une méthodologie pour évaluer efficacement des stratégies d'adaptation du step-size dans le cas de l'optimisation boîte-noire sans contraintes. Le step-size est un paramètre important dans les algorithmes évolutionnaires tels que les stratégies d'évolution; il contrôle la diversité de la population et, de ce fait, joue un rôle déterminant dans la convergence de l'algorithme. On présente aussi les résultats empiriques de la comparaison de trois méthodes d'adaptation du step-size. Ces algorithmes sont testés sur le testbed BBOB (black-box optimization benchmarking) de la plateforme COCO (comparing continuous optimisers). Dans la deuxième partie de cette thèse, sont présentées nos contributions dans le domaine de l'optimisation boîte-noire avec contraintes. On analyse la convergence linéaire d'algorithmes stochastiques adaptatifs pour l'optimisation sous contraintes dans le cas de contraintes linéaires, gérées avec une approche Lagrangien augmenté adaptative. Pour ce faire, on étend l'analyse par chaines de Markov faite dans le cas d'optimisation sans contraintes au cas avec contraintes: pour chaque algorithme étudié, on exhibe une classe de fonctions pour laquelle il existe une chaine de Markov homogène telle que la stabilité de cette dernière implique la convergence linéaire de l'algorithme. La convergence linéaire est déduite en appliquant une loi des grands nombres pour les chaines de Markov, sous l'hypothèse de la stabilité. Dans notre cas, la stabilité est validée empiriquement. / We investigate various aspects of adaptive randomized (or stochastic) algorithms for both constrained and unconstrained black-box continuous optimization. The first part of this thesis focuses on step-size adaptation in unconstrained optimization. We first present a methodology for assessing efficiently a step-size adaptation mechanism that consists in testing a given algorithm on a minimal set of functions, each reflecting a particular difficulty that an efficient step-size adaptation algorithm should overcome. We then benchmark two step-size adaptation mechanisms on the well-known BBOB noiseless testbed and compare their performance to the one of the state-of-the-art evolution strategy (ES), CMA-ES, with cumulative step-size adaptation. In the second part of this thesis, we investigate linear convergence of a (1 + 1)-ES and a general step-size adaptive randomized algorithm on a linearly constrained optimization problem, where an adaptive augmented Lagrangian approach is used to handle the constraints. To that end, we extend the Markov chain approach used to analyze randomized algorithms for unconstrained optimization to the constrained case. We prove that when the augmented Lagrangian associated to the problem, centered at the optimum and the corresponding Lagrange multipliers, is positive homogeneous of degree 2, then for algorithms enjoying some invariance properties, there exists an underlying homogeneous Markov chain whose stability (typically positivity and Harris-recurrence) leads to linear convergence to both the optimum and the corresponding Lagrange multipliers. We deduce linear convergence under the aforementioned stability assumptions by applying a law of large numbers for Markov chains. We also present a general framework to design an augmented-Lagrangian-based adaptive randomized algorithm for constrained optimization, from an adaptive randomized algorithm for unconstrained optimization.

A framework for optimization of reinsurance strategy is proposed for an insurance company with several lines of business (LoB), maximizing the Economic Value of purchasing reinsurance. The economic value is defined as the sum of the average ceded loss, the deducted risk premium, and the reduction in the cost of capital. The framework relies on simulated large claims per LoB rather than specific distributions, which gives more degrees of freedom to the insurance company.  Three models are presented, two non non-linear optimization models and a benchmark model. One non-linear optimization model is on individual LoB level and the other one is on company level with additional constraints using space bounded black box algorithms. The benchmark model is a Brute Force method using quantile discretization of potential retention levels, that helps to visualize the optimization surface.  The best results are obtained by a two-stage optimization using a mixture of global and local optimization algorithms. The economic value is maximized by 30% and reinsurance premium is halved if the optimization is made at the company level, by putting more emphasis on reduction in the cost of capital and less to average ceded loss. The results indicate an over-fitting when using VaR as the risk measure, impacting reduction in the cost of capital. As an alternative, Average VaR is recommended being numerically more robust.