Machine learning for image classification is a hot topic and it is increasing in popularity. Therefore the aim of this study is to provide a better understanding of convolutional neural network hyperparameters by comparing the test accuracy of convolutional neural network models with different hyperparameter value configurations. The focus of this study is to see whether there is an influence in the learning process depending on which hyperparameter values were used. For conducting the experiments convolutional neural network models were developed using the programming language Python utilizing the library Keras. The dataset used for this study iscifar-10, it includes 60000 colour images of 10 categories ranging from man-made objects to different animal species. Grid search is used for instantiating models with varying learning rate and momentum, width and depth values. Learning rate is only tested combined with momentum and width is only tested combined with depth. Activation functions, convolutional layers and batch size are tested individually. Grid search is compared against Bayesian optimization to see which technique will find the most optimized learning rate and momentum values. Results illustrate that the impact different hyperparameters have on the overall test accuracy varies. Learning rate and momentum affects the test accuracy greatly, however suboptimal values for learning rate and momentum can decrease the test accuracy severely. Activation function, width and depth, convolutional layer and batch size have a lesser impact on test accuracy. Regarding Bayesian optimization compared to grid search, results show that Bayesian optimization will not necessarily find more optimal hyperparameter values.

<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>

[pt] Os risers são importantes componentes na produção e exploração de petróleo e derivados. São responsáveis pelo transporte do óleo e gás encontrados no reservatório até a Unidade Estacionária de Produção (UEP) ou pela injeção de gás ou água no reservatório. A crescente demanda por esse produto faz com que a exploração seja feita em regiões com condições cada vez mais adversas. Tipicamente, um projeto deste porte exige um número muito grande de análises numéricas de elementos finitos e exigem uma experiência grande por parte do projetista a fim de obter uma solução viável. Esse desafio leva engenheiros a buscarem ferramentas consistentes e seguras que auxiliem nas etapas iniciais do projeto das configurações de risers e que sejam capazes de diminuir o número de análises totais exigidas. Uma dessas ferramentas é a utilização de métodos de otimização para obter de maneira consistente e segura os parâmetros que definem uma configuração. Este trabalho apresenta o método de Otimização Bayesiana, um método baseado em técnicas de aprendizado de máquina capaz de resolver problemas de otimização do tipo caixa-preta de maneira eficiente explorando o uso de aproximações analíticas da função objetivo, que se quer otimizar. O método é aplicado em diferentes estudos de casos visando validálo como capaz de resolver problemas de configuração de riser de maneira eficiente e consistente. Dentre os problemas aplicados estão diferentes tipos de configurações, diferentes casos realistas, mono-objetivo e multi-objetivo. / [en] Risers are an important component in the oil s production and exploration field. They are responsible for the oil and gas transportation from the reservoir to the floating unit or injection of gas or water into the reservoir. The increasing the demand for this product has lead projects to explore to areas in which conditions are harsher. Typically, such a large project demands a large number of numerical finite element analyses and a great expertise from the engineer in charge in order to obtain a viable solution. This challenge leads engineers in search of consistent and reliable tools that assist in the early stages of the riser configuration design and are capable of reducing the number of total analyses required. One of these tools is application of optimization methods to obtain in a consistent and reliable manner the parameters which define a configuration. This work presents the Bayesian Optimization method, a method based on machine learning techniques capable of efficiently solving so called black box problems by exploring analytical approximations of the objective function, the function to be minimized. The method is applied to different case studies aiming to validate it as capable of solving a wide variety of riser configuration problems in an efficient and consistent way. Among the problems applied are different types of configurations, different realistic cases, mono-objective and multi-objective.

L’apprentissage par renforcement profond a mené à de nombreux résultats prometteurs dans l’apprentissage des jeux vidéo à partir de pixels, dans la robotique pour l’apprentissage de compétences généralisables et dans les soins de santé pour l’apprentissage de traitement dynamiques. Un obstacle demeure toutefois: celui du manque d’efficacité dans le nombre d’échantillons nécessaires pour obtenir de bons résultats. Pour résoudre ce problème, notre objectif est d’améliorer l’efficacité de l’apprentissage en améliorant les capacité d’acquisition de nouvelles données, un problème d’exploration. L’approche proposée consiste à : (1) Apprendre un ensemble diversifié d’environments (donnant lieu à un changement de dynamique) (2) Apprendre une politique capable de mieux s’adapter aux changements dans l’envi- ronnement, à l’aide du méta-apprentissage. Cette méthode peut avoir des impacts bénéfiques dans de nombreux problèmes du monde réel tels que la découverte de médicaments, dans laquelle nous sommes confrontés à un espace d’actions très grand. D’autant plus, la conception de nouvelles substances thérapeutiques qui sont fonctionnellement intéressantes nécessite une exploration efficace du paysage de la recherche. / Deep reinforcement learning has led to promising results in learning video games from pixels, robotics for learning generalizable skills, and healthcare for learning dynamic treatments. However, an obstacle remains the lack of efficiency in the number of samples required to achieve good results. To address this problem, our goal is to improve sample efficiency by improving the ability to acquire new data, an issue of exploration. The proposed approach is to: (1) Learn a diverse set of environments (resulting in a change of dynamics) (2) earn a policy that can better adapt to changes in the environment using meta-learning This method can benefit many real-world problems, such as drug discovery, where we face a large action space. Furthermore, designing new therapeutic substances that are functionally interesting requires efficient exploration of the research landscape

In the fields of process control engineering and robotics, especially in automatic control, optimization challenges frequently manifest as complex problems with expensive evaluations. This thesis zeroes in on one such problem: the optimization of camera positions for Convolutional Neural Networks (CNNs). CNNs have specific attention points in images that are often not intuitive to human perception, making camera placement critical for performance. The research is guided by two primary questions. The first investigates the role of Explainable Artificial Intelligence (XAI), specifically GradCAM++ visual explanations, in Computer Vision for aiding in the evaluation of different camera positions. Building on this, the second question assesses a novel algorithm that leverages these XAI features against traditional black-box optimization methods. To answer these questions, the study employs a robotic auto-positioning system for data collection, CNN model training, and performance evaluation. A case study focused on classifying flow regimes in industrial-grade bioreactors validates the method. The proposed approach shows improvements over established techniques like Grid Search, Random Search, Bayesian optimization, and Simulated Annealing. Future work will focus on gathering more data and including noise for generalized conclusions.:Contents 1 Introduction 1.1 Motivation 1.2 Problem Analysis 1.3 Research Question 1.4 Structure of the Thesis 2 State of the Art 2.1 Literature Research Methodology 2.1.1 Search Strategy 2.1.2 Inclusion and Exclusion Criteria 2.2 Blackbox Optimization 2.3 Mathematical Notation 2.4 Bayesian Optimization 2.5 Simulated Annealing 2.6 Random Search 2.7 Gridsearch 2.8 Explainable A.I. and Saliency Maps 2.9 Flowregime Classification in Stirred Vessels 2.10 Performance Metrics 2.10.1 R2 Score and Polynomial Regression for Experiment Data Analysis 2.10.2 Blackbox Optimization Performance Metrics 2.10.3 CNN Performance Metrics 3 Methodology 3.1 Requirement Analysis and Research Hypothesis 3.2 Research Approach: Case Study 3.3 Data Collection 3.4 Evaluation and Justification 4 Concept 4.1 System Overview 4.2 Data Flow 4.3 Experimental Setup 4.4 Optimization Challenges and Approaches 5 Data Collection and Experimental Setup 5.1 Hardware Components 5.2 Data Recording and Design of Experiments 5.3 Data Collection 5.4 Post-Experiment 6 Implementation 6.1 Simulation Unit 6.2 Recommendation Scalar from Saliency Maps 6.3 Saliency Map Features as Guidance Mechanism 6.4 GradCam++ Enhanced Bayesian Optimization 6.5 Benchmarking Unit 6.6 Benchmarking 7 Results and Evaluation 7.1 Experiment Data Analysis 7.2 Recommendation Scalar 7.3 Benchmarking Results and Quantitative Analysis 7.3.1 Accuracy Results from the Benchmarking Process 7.3.2 Cumulative Results Interpretation 7.3.3 Analysis of Variability 7.4 Answering the Research Questions 7.5 Summary 8 Discussion 8.1 Critical Examination of Limitations 8.2 Discussion of Solutions to Limitations 8.3 Practice-Oriented Discussion of Findings 9 Summary and Outlook / Im Bereich der Prozessleittechnik und Robotik, speziell bei der automatischen Steuerung, treten oft komplexe Optimierungsprobleme auf. Diese Arbeit konzentriert sich auf die Optimierung der Kameraplatzierung in Anwendungen, die Convolutional Neural Networks (CNNs) verwenden. Da CNNs spezifische, für den Menschen nicht immer ersichtliche, Merkmale in Bildern hervorheben, ist die intuitive Platzierung der Kamera oft nicht optimal. Zwei Forschungsfragen leiten diese Arbeit: Die erste Frage untersucht die Rolle von Erklärbarer Künstlicher Intelligenz (XAI) in der Computer Vision zur Bereitstellung von Merkmalen für die Bewertung von Kamerapositionen. Die zweite Frage vergleicht einen darauf basierenden Algorithmus mit anderen Blackbox-Optimierungstechniken. Ein robotisches Auto-Positionierungssystem wird zur Datenerfassung und für Experimente eingesetzt. Als Lösungsansatz wird eine Methode vorgestellt, die XAI-Merkmale, insbesondere solche aus GradCAM++ Erkenntnissen, mit einem Bayesschen Optimierungsalgorithmus kombiniert. Diese Methode wird in einer Fallstudie zur Klassifizierung von Strömungsregimen in industriellen Bioreaktoren angewendet und zeigt eine gesteigerte performance im Vergleich zu etablierten Methoden. Zukünftige Forschung wird sich auf die Sammlung weiterer Daten, die Inklusion von verrauschten Daten und die Konsultation von Experten für eine kostengünstigere Implementierung konzentrieren.:Contents 1 Introduction 1.1 Motivation 1.2 Problem Analysis 1.3 Research Question 1.4 Structure of the Thesis 2 State of the Art 2.1 Literature Research Methodology 2.1.1 Search Strategy 2.1.2 Inclusion and Exclusion Criteria 2.2 Blackbox Optimization 2.3 Mathematical Notation 2.4 Bayesian Optimization 2.5 Simulated Annealing 2.6 Random Search 2.7 Gridsearch 2.8 Explainable A.I. and Saliency Maps 2.9 Flowregime Classification in Stirred Vessels 2.10 Performance Metrics 2.10.1 R2 Score and Polynomial Regression for Experiment Data Analysis 2.10.2 Blackbox Optimization Performance Metrics 2.10.3 CNN Performance Metrics 3 Methodology 3.1 Requirement Analysis and Research Hypothesis 3.2 Research Approach: Case Study 3.3 Data Collection 3.4 Evaluation and Justification 4 Concept 4.1 System Overview 4.2 Data Flow 4.3 Experimental Setup 4.4 Optimization Challenges and Approaches 5 Data Collection and Experimental Setup 5.1 Hardware Components 5.2 Data Recording and Design of Experiments 5.3 Data Collection 5.4 Post-Experiment 6 Implementation 6.1 Simulation Unit 6.2 Recommendation Scalar from Saliency Maps 6.3 Saliency Map Features as Guidance Mechanism 6.4 GradCam++ Enhanced Bayesian Optimization 6.5 Benchmarking Unit 6.6 Benchmarking 7 Results and Evaluation 7.1 Experiment Data Analysis 7.2 Recommendation Scalar 7.3 Benchmarking Results and Quantitative Analysis 7.3.1 Accuracy Results from the Benchmarking Process 7.3.2 Cumulative Results Interpretation 7.3.3 Analysis of Variability 7.4 Answering the Research Questions 7.5 Summary 8 Discussion 8.1 Critical Examination of Limitations 8.2 Discussion of Solutions to Limitations 8.3 Practice-Oriented Discussion of Findings 9 Summary and Outlook

No description available.

Notwithstanding the recent successes of deep convolutional neural networks for classification tasks, they are sensitive to the selection of their hyperparameters, which impose an exponentially large search space on modern convolutional models. Traditional hyperparameter selection methods include manual, grid, or random search, but these require expert knowledge or are computationally burdensome. Divergently, Bayesian optimization and evolutionary inspired techniques have surfaced as viable alternatives to the hyperparameter problem. Thus, an alternative hybrid approach that combines the advantages of these techniques is proposed. Specifically, the search space is partitioned into discrete-architectural, and continuous and categorical hyperparameter subspaces, which are respectively traversed by a stochastic genetic search, followed by a genetic-Bayesian search. Simulations on a prominent image classification task reveal that the proposed method results in an overall classification accuracy improvement of 0.87% over unoptimized baselines, and a greater than 97% reduction in computational costs compared to a commonly employed brute force approach. / Electrical and Mining Engineering / M. Tech. (Electrical Engineering)

Le réservoir de carbone de sol est un élément clé du cycle global du carbone et donc du système climatique. Les sols et le carbone organique qu'ils contiennent constituent le plus grand réservoir de carbone des écosystèmes terrestres. Ce réservoir est également responsable du stockage d'une grande quantité de carbone prélevé de l'atmosphère par les plantes par la photosynthèse. C'est pourquoi les sols sont considérés comme une stratégie de mitigation viable pour réduire la concentration atmosphérique de CO2 dûe aux émissions globales de CO2 d'origine fossile. Malgré son importance, des incertitudes subsistent quant à la taille du réservoir global de carbone organique de sol et à ses dynamiques. Les modèles de biosphère terrestre sont des outils essentiels pour quantifier et étudier la dynamique du carbone organique de sol. Ces modèles simulent les processus biophysiques et biogéochimiques au sein des écosystèmes et peuvent également simuler le comportement futur du réservoir de carbone organique de sol en utilisant des forçages météorologiques appropriés. Cependant, de grandes incertitudes dans les projections faite par les modèles de biosphère terrestre sur les dynamiques du carbone organique de sol ont été observées, en partie dues au problème de l'équifinalité. Afin d'améliorer notre compréhension de la dynamique du carbone organique de sol, cette recherche visait à optimiser les paramètres du schéma de carbone de sol contenu dans le modèle de schéma canadien de surface terrestre incluant les cycles biogéochimiques (CLASSIC), afin de parvenir à une meilleure représentation de la dynamique du carbone organique de sol. Une analyse de sensibilité globale a été réalisée pour identifier lesquels parmis les 16 paramètres du schéma de carbone de sol, n'affectaient pas la simulation du carbone organique de sol et de la respiration du sol. L'analyse de sensibilité a utilisé trois sites de covariance des turbulences afin de représenter différentes conditions climatiques simulées par le schéma de carbone de sol et d'économiser le coût calculatoire de l'analyse. L'analyse de sensibilité a démontré que certains paramètres du schéma de carbone de sol ne contribuent pas à la variance des simulations du carbone organique de sol et de la respiration du sol. Ce résultat a permis de réduire la dimensionnalité du problème d'optimisation. Ensuite, quatre scénarios d'optimisation ont été élaborés sur la base de l'analyse de sensibilité, chacun utilisant un ensemble de paramètres. Deux fonctions coûts ont été utilisées pour l'optimisation de chacun des scénarios. L'optimisation a également démontré que la fonction coût utilisée avait un impact sur les ensembles de paramètres optimisés. Les ensembles de paramètres obtenus à partir des différents scénarios et fonctions coûts ont été comparés à des ensembles de données indépendants et à des estimations globales du carbone organique de sol à l'aide de métrique tel la racine de l'erreur quadratique moyenne et le bias, afin d'évaluer l'effet des ensembles de paramètres sur les simulations effectuées par le schéma de carbone de sol. Un ensemble de paramètres a surpassé les autres ensembles de paramètres optimisés ainsi que le paramétrage par défaut du modèle. Ce résultat a indiqué que la structure d'optimisation était en mesure de produire un ensemble de paramètres qui simulait des valeurs de carbone organique de sol et de respiration du sol qui étaient plus près des valeurs observées que le modèle CLASSIC par défaut, améliorant la représentation de la dynamique du carbone du sol. Cet ensemble de paramètres optimisés a ensuite été utilisé pour effectuer des simulations futures (2015-2100) de la dynamique du carbone organique de sol afin d'évaluer son impact sur les projections de CLASSIC. Les simulations futures ont montré que l'ensemble de paramètres optimisés simulait une quantité de carbone organique de sol 62 % plus élevée que l'ensemble de paramètres par défaut tout en simulant des flux de respiration du sol similaires. Les simulations futures ont également montré que les ensembles de paramètres optimisés et par défaut prévoyaient que le réservoir de carbone organique de sol demeurerait un puits de carbone net d'ici 2100 avec des sources nettes régionales. Cette étude a amélioré globalement la représentation de la dynamique du carbone organique de sol dans le schéma de carbone de sol de CLASSIC en fournissant un ensemble de paramètres optimisés. Cet ensemble de paramètres devrait permettre d'améliorer notre compréhension de la dynamique du carbone du sol. / The soil carbon pool is a vital component of the global carbon cycle and, therefore, the climate system. Soil organic carbon (SOC) is the largest carbon pool in terrestrial ecosystems. This pool stores a large quantity of carbon that plants have removed from the atmosphere through photosynthesis. Because of this, soils are considered a viable climate change mitigation strategy to lower the global atmospheric CO2 concentration that is presently being driven higher by anthropogenic fossil CO2 emissions. Despite its importance, there are still considerable uncertainties around the size of the global SOC pool and its response to changing climate. Terrestrial biosphere models (TBM) simulate the biogeochemical processes within ecosystems and are critical tools to quantify and study SOC dynamics. These models can also simulate the future behavior of SOC if carefully applied and given the proper meteorological forcings. However, TBM predictions of SOC dynamics have high uncertainties due in part to equifinality. To improve our understanding of SOC dynamics, this research optimized the parameters of the soil carbon scheme contained within the Canadian Land Surface Scheme Including Biogeochemical Cycles (CLASSIC), to better represent SOC dynamics. A global sensitivity analysis was performed to identify which of the 16 parameters of the soil carbon scheme did not affect simulated SOC stocks and soil respiration (Rsoil). The sensitivity analysis used observations from three eddy covariance sites for computational efficiency and to encapsulate the climate represented by the global soil carbon scheme. The sensitivity analysis revealed that some parameters of the soil carbon scheme did not contribute to the variance of simulated SOC and Rsoil. These parameters were excluded from the optimization which helped reduce the dimensionality of the optimization problem. Then, four optimization scenarios were created based on the sensitivity analysis, each using a different set of parameters to assess the impact the number of parameters included had on the optimization. Two different loss functions were used in the optimization to assess the impact of accounting for observational error. Comparing the optimal parameters between the optimizations performed using the different loss functions showed that the loss functions impacted the optimized parameter sets. To determine which optimized parameter set obtained by each loss function was most skillful, they were compared to independent data sets and global estimates of SOC, which were not used in the optimization using comparison metrics based on root-mean-square-deviation and bias. This study generated an optimal parameter set that outperformed the default parameterization of the model. This optimal parameter set was then applied in future simulations of SOC dynamics to assess its impact upon CLASSIC's future projections. These future simulations showed that the optimal parameter set simulated future global SOC content 62 % higher than the default parameter set while simulating similar Rsoil fluxes. The future simulations also showed that both the optimized and default parameter sets projected that the SOC pool would be a net sink by 2100 with regional net sources, notably tropical regions.

Side-channel attacks have become a realistic threat to the implementations of cryptographic algorithms. By analyzing the unintentional, side-channel leakage, the attacker is able to recover the secret of the target. Recently, a new type of side-channel leakage has been discovered, called far field EM emissions. Unlike attacks based on near field EM emissions or power consumption, the attack based on far field EM emissions is able to extract the secret key from the victim device of several meters distance. However, existing deep-learning attacks based far field EM commonly use a random or grid search method to optimize neural networks’ hyperparameters. Recently, an automated way for deep learning hyperparameter tuning based on Auto- Keras library, called AutoSCA framework, was applied to near-field EM attacks. In this work, we investigate if AutoSCA could help far field EM side-channel attacks. In our experiments, the target is a Bluetooth-5 supported Nordic Semiconductor nRF52832 development kit implementation of Advanced Encryption Standard (AES). Our experiments show that, by using a deep-learning model generated by the AutoSCA framework, we need 485 traces on average to recover a subkey from traces captured at 15 meters distance from the victim device without repeating each encryption. For the same conditions, the state-of-the-art method uses 510 traces. Furthermore, our model contains only 667,433 trainable parameters in total, implying that it requires roughly 9 times less training resources compared to the larger models in the previous work. / Angrepp på sidokanaler har blivit ett realistiskt hot mot implementeringen av kryptografiska algoritmer.Genom att analysera det oavsiktliga läckaget kan angriparen hitta hemligheten bakom målet.Nyligen har en ny typ av sidokanalläckage upptäckts, kallad fjärrfälts EM-utsläpp.Till skillnad från attacker baserade på nära fält EM- utsläpp eller energiförbrukning, kan attacken baserad på yttre fält EM-utsläpp extrahera den hemliga nyckeln från den skadade anordningen på flera meter avstånd.Men befintliga djupinlärningsattacker baserade på långt fält EM använder ofta en slumpmässig sökmetod för att optimera nervnätens hyperparametrar. Nyligen tillämpades ett automatiserat sätt för djupinlärning av hyperparametern baserad på Auto-Keras- bibliotek, AutoSCA- ramverket, vid EM-angrepp nära fältet.I det här arbetet undersöker vi om AutoSCA kan hjälpa till med EM-angrepp.I våra experiment är målet en Bluetooth-5-stödd nordisk semidirigent nR52832- utvecklingsutrustning för avancerad krypteringsstandard (AES).Våra experiment visar att genom att använda en djupinlärningsmodell skapad av AutoSCA-ramverket, behöver vi 485-spår i genomsnitt för att hämta en subnyckel från spår tagna på 15- meters avstånd från offrets apparat utan att upprepa varje kryptering.Under samma förhållanden använder den senaste metoden 510-spår.Dessutom innehåller vår modell bara 667,433-parametrar som totalt kan användas för utbildning, vilket innebär att det krävs ungefär nio gånger mindre utbildningsresurser jämfört med de större modellerna i det tidigare arbetet.

Choosing the correct angle of electrical tilt in a radio base station is essential when optimizing for coverage and capacity. A reinforcement learning agent can be trained to make this choice. If the training of the agent in the real world is restricted or even impossible, alternative methods can be used. Training in simulation combined with an approximation of the real world is one option that comes with a set of challenges associated with the reality gap. In this thesis, a method based on Bayesian optimization is implemented to tune the environment in which domain randomization is performed to improve the quality of the simulation training. The results show that using Bayesian optimization to find a good subset of parameters works even when access to the real world is constrained. Two off- policy estimators based on inverse propensity scoring and direct method evaluation in combination with an offline dataset of previously collected cell traces were tested. The method manages to find an isolated subspace of the whole domain that optimizes the randomization while still giving good performance in the target domain. / Rätt val av elektrisk antennvinkel för en radiobasstation är avgörande när täckning och kapacitetsoptimering (eng. coverage and capacity optimization) görs för en förstärkningsinlärningsagent. Om träning av agenten i verkligheten är besvärlig eller till och med omöjlig att genomföra kan olika alternativa metoder användas. Simuleringsträning kombinerad med en skattningsmodell av verkligheten är ett alternativ som har olika utmaningar kopplade till klyftan mellan simulering och verkligheten (eng. reality gap). I denna avhandling implementeras en lösning baserad på Bayesiansk Optimering med syftet att anpassa miljön som domänrandomisering sker i för att förbättra kvaliteten på simuleringsträningen. Resultatet visar att Bayesiansk Optimering kan användas för att hitta ett urval av fungerande parametrar även när tillgången till den faktiska verkligheten är begränsad. Två skattningsmodeller baserade på invers propensitetsviktning och direktmetodutvärdering i kombination med ett tidigare insamlat dataset av nätverksdata testades. Den tillämpade metoden lyckas hitta ett isolerat delrum av parameterrymden som optimerar randomiseringen samtidigt som prestationen i verkligheten hålls på en god nivå.