Reinforcement learning is a machine learning technique in which an artificial intelligence agent is guided by positive and negative rewards to learn strategies. To guide the agent’s behavior in addition to the reward are its hyperparameters. These values control how the agent learns. These hyperparameters are rarely disclosed in contemporary research, making it hard to estimate the value of optimizing these hyperparameters. This study aims to partly compare three different popular reinforcement learning algorithms, Proximal Policy Optimization (PPO), Advantage Actor-Critic (A2C) and Deep Q Network (DQN), and partly investigate the effects of hyperparameter optimization of several hyperparameters for each algorithm. All the included algorithms showed a significant difference after hyperparameter optimization, resulting in higher performance. A2C showed the largest performance increase after hyperparameter optimization, and PPO performed the best of the three algorithms both with default and optimized hyperparameters.

Affective computing has gained significant attention from researchers in the last decade due to the wide variety of applications that can benefit from this technology. Often, researchers describe affect using emotional dimensions such as arousal and valence. Valence refers to the spectrum of negative to positive emotions while arousal determines the level of excitement. Describing emotions through continuous dimensions (e.g. valence and arousal) allows us to encode subtle and complex affects as opposed to discrete emotions, such as the basic six emotions: happy, anger, fear, disgust, sad and neutral. Recognizing spontaneous and subtle emotions remains a challenging problem for computers. In our work, we employ two modalities of information: video and audio. Hence, we extract visual and audio features using deep neural network models. Given that emotions are time-dependent, we apply the Temporal Convolutional Neural Network (TCN) to model the variations in emotions. Additionally, we investigate an alternative model that combines a Convolutional Neural Network (CNN) and a Recurrent Neural Network (RNN). Given our inability to fit the latter deep model into the main memory, we divide the RNN into smaller segments and propose a scheme to back-propagate gradients across all segments. We configure the hyperparameters of all models using Gaussian processes to obtain a fair comparison between the proposed models. Our results show that TCN outperforms RNN for the recognition of the arousal and valence emotional dimensions. Therefore, we propose the adoption of TCN for emotion detection problems as a baseline method for future work. Our experimental results show that TCN outperforms all RNN based models yielding a concordance correlation coefficient of 0.7895 (vs. 0.7544) on valence and 0.8207 (vs. 0.7357) on arousal on the validation dataset of SEWA dataset for emotion prediction.

Machine learning and its wide range of applications is becoming increasingly prevalent in both academia and industry. This thesis will focus on the two machine learning methods convolutional neural networks and reinforcement learning. Convolutional neural networks has seen great success in various applications for both classification and regression problems in a diverse range of fields, e.g. vision for self-driving cars or facial recognition. These networks are built on a set of trainable weights optimized on data, and a set of hyperparameters set by the designer of the network which will remain constant. For the network to perform well, the hyperparameters have to be optimized separately. The goal of this thesis is to investigate the use of reinforcement learning as a method for optimizing hyperparameters in convolutional neural networks built for classification problems. The reinforcement learning methods used are a tabular Q-learning and a new Q-learning inspired algorithm denominated max-table. These algorithms have been tested with different exploration policies based on each hyperparameter value’s covariance, precision or relevance to the performance metric. The reinforcement learning algorithms were mostly tested on the datasets CIFAR10 and MNIST fashion against a baseline set by random search. While the Q-learning algorithm was not able to perform better than random search, max-table was able to perform better than random search in 50% of the time on both datasets. Hyperparameterbased exploration policy using covariance and relevance were shown to decrease the optimizers’ performance. No significant difference was found between a hyperparameter based exploration policy using performance and an equally distributed exploration policy. / Maskininlärning och dess många tillämpningsområden blir vanligare i både akademin och industrin. Den här uppsatsen fokuserar på två maskininlärningsmetoder, faltande neurala nätverk och förstärkningsinlärning. Faltande neurala nätverk har sett stora framgångar inom olika applikationsområden både för klassifieringsproblem och regressionsproblem inom diverse fält, t.ex. syn för självkörande bilar eller ansiktsigenkänning. Dessa nätverk är uppbyggda på en uppsättning av tränbara parameterar men optimeras på data, samt en uppsättning hyperparameterar bestämda av designern och som hålls konstanta vilka behöver optimeras separat för att nätverket ska prestera bra. Målet med denna uppsats är att utforska användandet av förstärkningsinlärning som en metod för att optimera hyperparameterar i faltande neurala nätverk gjorda för klassifieringsproblem. De förstärkningsinlärningsmetoder som använts är en tabellarisk "Q-learning" samt en ny "Q-learning" inspirerad metod benämnd "max-table". Dessa algoritmer har testats med olika handlingsmetoder för utforskning baserade på hyperparameterarnas värdens kovarians, precision eller relevans gentemot utvärderingsmetriken. Förstärkningsinlärningsalgoritmerna var i största del testade på dataseten CIFAR10 och MNIST fashion och jämförda mot en baslinje satt av en slumpmässig sökning. Medan "Q-learning"-algoritmen inte kunde visas prestera bättre än den slumpmässiga sökningen, kunde "max-table" prestera bättre på 50\% av tiden på både dataseten. De handlingsmetoder för utforskning som var baserade på kovarians eller relevans visades minska algoritmens prestanda. Ingen signifikant skillnad kunde påvisas mellan en handlingsmetod baserad på hyperparametrarnas precision och en jämnt fördelad handlingsmetod för utforsking.

No description available.

Maskininlärning har blivit allt vanligare inom näringslivet. Informationsinsamling med Data mining (DM) har expanderats och DM-utövare använder en mängd tumregler för att effektivisera tillvägagångssättet genom att undvika en anständig tid att ställa in hyperparametrarna för en given ML-algoritm för nå bästa träffsäkerhet. Förslaget i denna rapport är att införa ett tillvägagångssätt som systematiskt optimerar ML-algoritmerna med hjälp av genetiska algoritmer (GA), utvärderar om och hur modellen ska konstrueras för att hitta globala lösningar för en specifik datamängd. Genom att implementera genetiska algoritmer på två utvalda ML-algoritmer, K-nearest neighbors och Random forest, med två numeriska datamängder, Iris-datauppsättning och Wisconsin-bröstcancerdatamängd. Modellen utvärderas med träffsäkerhet och beräkningstid som sedan jämförs med sökmetoden exhaustive search. Resultatet har visat att GA fungerar bra för att hitta bra träffsäkerhetspoäng på en rimlig tid. Det finns vissa begränsningar eftersom parameterns betydelse varierar för olika ML-algoritmer. / As machine learning (ML) is being more and more frequent in the business world, information gathering through Data mining (DM) is on the rise, and DM-practitioners are generally using several thumb rules to avoid having to spend a decent amount of time to tune the hyperparameters (parameters that control the learning process) of an ML algorithm to gain a high accuracy score. The proposal in this report is to conduct an approach that systematically optimizes the ML algorithms using genetic algorithms (GA) and to evaluate if and how the model should be constructed to find global solutions for a specific data set. By implementing a GA approach on two ML-algorithms, K-nearest neighbors, and Random Forest, on two numerical data sets, Iris data set and Wisconsin breast cancer data set, the model is evaluated by its accuracy scores as well as the computational time which then is compared towards a search method, specifically exhaustive search. The results have shown that it is assumed that GA works well in finding great accuracy scores in a reasonable amount of time. There are some limitations as the parameter’s significance towards an ML algorithm may vary.

La récente révolution de l'apprentissage automatique s'est fortement appuyée sur l'utilisation de bancs de test standardisés. Ces derniers sont au centre de la méthodologie scientifique en apprentissage automatique, fournissant des cibles et mesures indéniables des améliorations des algorithmes d'apprentissage. Ils ne garantissent cependant pas la validité des résultats ce qui implique que certaines conclusions scientifiques sur les avancées en intelligence artificielle peuvent s'avérer erronées. Nous abordons cette question dans cette thèse en soulevant d'abord la problématique (Chapitre 5), que nous étudions ensuite plus en profondeur pour apporter des solutions (Chapitre 6) et finalement developpons un nouvel outil afin d'amélioration la méthodologie des chercheurs (Chapitre 7). Dans le premier article, chapitre 5, nous démontrons la problématique de la reproductibilité pour des bancs de test stables et consensuels, impliquant que ces problèmes sont endémiques aussi à de grands ensembles d'applications en apprentissage automatique possiblement moins stable et moins consensuels. Dans cet article, nous mettons en évidence l'impact important de la stochasticité des bancs de test, et ce même pour les plus stables tels que la classification d'images. Nous soutenons d'après ces résultats que les solutions doivent tenir compte de cette stochasticité pour améliorer la reproductibilité des bancs de test. Dans le deuxième article, chapitre 6, nous étudions les différentes sources de variation typiques aux bancs de test en apprentissage automatique, mesurons l'effet de ces variations sur les méthodes de comparaison d'algorithmes et fournissons des recommandations sur la base de nos résultats. Une contribution importante de ce travail est la mesure de la fiabilité d'estimateurs peu coûteux à calculer mais biaisés servant à estimer la performance moyenne des algorithmes. Tel qu'expliqué dans l'article, un estimateur idéal implique plusieurs exécution d'optimisation d'hyperparamètres ce qui le rend trop coûteux à calculer. La plupart des chercheurs doivent donc recourir à l'alternative biaisée, mais nous ne savions pas jusqu'à présent la magnitude de la dégradation de cet estimateur. Sur la base de nos résultats, nous fournissons des recommandations pour la comparison d'algorithmes sur des bancs de test avec des budgets de calculs limités. Premièrement, les sources de variations devraient être randomisé autant que possible. Deuxièmement, la randomization devrait inclure le partitionnement aléatoire des données pour les ensembles d'entraînement, de validation et de test, qui s'avère être la plus importante des sources de variance. Troisièmement, des tests statistiques tel que la version du Mann-Withney U-test présenté dans notre article devrait être utilisé plutôt que des comparisons sur la simple base de moyennes afin de prendre en considération l'incertitude des mesures de performance. Dans le chapitre 7, nous présentons un cadriciel d'optimisation d'hyperparamètres développé avec principal objectif de favoriser les bonnes pratiques d'optimisation des hyperparamètres. Le cadriciel est conçu de façon à privilégier une interface simple et intuitive adaptée aux habitudes de travail des chercheurs en apprentissage automatique. Il inclut un nouveau système de versionnage d'expériences afin d'aider les chercheurs à organiser leurs itérations expérimentales et tirer profit des résultats antérieurs pour augmenter l'efficacité de l'optimisation des hyperparamètres. L'optimisation des hyperparamètres joue un rôle important dans les bancs de test, les hyperparamètres étant un facteur confondant significatif. Fournir aux chercheurs un instrument afin de bien contrôler ces facteurs confondants est complémentaire aux recommandations pour tenir compte des sources de variation dans le chapitre 6. Nos recommendations et l'outil pour l'optimisation d'hyperparametre offre une base solide pour une méthodologie robuste et fiable. / The recent revolution in machine learning has been strongly based on the use of standardized benchmarks. Providing clear target metrics and undeniable measures of improvements of learning algorithms, they are at the center of the scientific methodology in machine learning. They do not ensure validity of results however, therefore some scientific conclusions based on flawed methodology may prove to be wrong. In this thesis we address this question by first raising the issue (Chapter 5), then we study it to find solutions and recommendations (Chapter 6) and build tools to help improve the methodology of researchers (Chapter 7). In first article, Chapter 5, we demonstrate the issue of reproducibility in stable and consensual benchmarks, implying that these issues are endemic to a large ensemble of machine learning applications that are possibly less stable or less consensual. We raise awareness of the important impact of stochasticity even in stable image classification tasks and contend that solutions for reproducible benchmarks should account for this stochasticity. In second article, Chapter 6, we study the different sources of variation that are typical in machine learning benchmarks, measure their effect on comparison methods to benchmark algorithms and provide recommendations based on our results. One important contribution of this work is that we measure the reliability of a cheaper but biased estimator for the average performance of algorithms. As explained in the article, an ideal estimator involving multiple rounds of hyperparameter optimization is too computationally expensive. Most researchers must resort to use the biased alternative, but it has been unknown until now how serious a degradation of the quality of estimation this leads to. Our investigations provides guidelines for benchmarks on practical budgets. First, as many sources of variations as possible should be randomized. Second, the partitioning of data in training, validation and test sets should be randomized as well, since this is the most important source of variation. Finally, statistical tests should be used instead of ad-hoc average comparisons so that the uncertainty of performance estimation can be accounted for when comparing machine learning algorithms. In Chapter 7, we present a framework for hyperparameter optimization that has been developed with the main goal of encouraging best practices for hyperparameter optimization. The framework is designed to favor a simple and intuitive interface adapted to the workflow of machine learning researchers. It includes a new version control system for experiments to help researchers organize their rounds of experimentations and leverage prior results for more efficient hyperparameter optimization. Hyperparameter optimization plays an important role in benchmarking, with the effect of hyperparameters being a serious confounding factor. Providing an instrument for researchers to properly control this confounding factor is complementary to our guidelines to account for sources of variation in Chapter 7. Our recommendations together with our tool for hyperparameter optimization provides a solid basis for a reliable methodology in machine learning benchmarks.

<p dir="ltr">Neural networks (NNs) effectively address the challenges of engineering design optimization by using data-driven models, thus reducing computational demands. However, their effectiveness depends heavily on hyperparameter optimization (HPO), which is a global optimization problem. While traditional HPO methods, such as manual, grid, and random search, are simple, they often fail to navigate the vast hyperparameter (HP) space efficiently. This work examines the effectiveness of integrating Bayesian optimization (BO) with multi-armed bandit (MAB) optimization for HPO in NNs. The thesis initially addresses HPO in one-shot sampling, where NNs are trained using datasets of varying sample sizes. It compares the performance of NNs optimized through traditional HPO techniques and a combination of BO and MAB optimization on the analytical Branin function and aerodynamic shape optimization (ASO) of an airfoil in transonic flow. Findings from the optimization of the Branin function indicate that the combined BO and MAB optimization approach leads to simpler NNs and reduces the sample size by approximately 10 to 20 compared to traditional HPO methods, all within half the time. This efficiency improvement is even more pronounced in ASO, where the BO and MAB optimization use about 100 fewer samples than the traditional methods to achieve the optimized airfoil design. The thesis then expands on adaptive HPOs within the framework of efficient global optimization (EGO) using a NN-based prediction and uncertainty (EGONN) algorithm. It employs the BO and MAB optimization for tuning HPs during sequential sampling, either every iteration (HPO-1itr) or every five iterations (HPO-5itr). These strategies are evaluated against the EGO as a benchmark method. Through experimentation with the analytical three-dimensional Hartmann function and ASO, assessing both comprehensive and selective tunable HP sets, the thesis contrasts adaptive HPO approaches with a static HPO method (HPO-static), which uses the initial HP settings throughout. Initially, a comprehensive set of the HPs is optimized and evaluated, followed by an examination of selectively chosen HPs. For the optimization of the three-dimensional Hartmann function, the adaptive HPO strategies surpass HPO-static in performance in both cases, achieving optimal convergence and sample efficiency comparable to EGO. In ASO, applying the adaptive HPO strategies reduces the baseline airfoil's drag coefficient to 123 drag counts (d.c.) for HPO-1itr and 120 d.c. for HPO-5itr when tuning the full set of the HPs. For a selected subset of the HPs, 123 d.c. and 121 d.c. are achieved by HPO-1itr and HPO-5itr, respectively, which are comparable to the minimum achieved by EGO. While the HPO-static method reduces the drag coefficient to 127 d.c. by tuning a subset of the HPs, which is a 15 d.c. reduction from its full set case, it falls short of the minimum of adaptive HPO strategies. Focusing on a subset of the HPs reduces time costs and enhances the convergence rate without sacrificing optimization efficiency. The time reduction is more significant with higher HPO frequencies as HPO-1itr cuts time by 66%, HPO-5itr by 38%, and HPO-static by 2%. However, HPO-5itr still requires 31% of the time needed by HPO-1itr for the full HP tuning and 56% for the subset HP tuning.</p>

Reservoir computers rely on an internal network to predict the future state(s) of dynamical processes. To understand how a reservoir's accuracy depends on this network, we study how varying the networ's topology and scaling affects the reservoir's ability to predict the chaotic dynamics on the Lorenz attractor. We define a metric for diversity, the property describing the variety of the responses of the nodes that make up reservoir's internal network. We use Bayesian hyperparameter optimization to find optimal hyperparameters and explore the relationships between diversity, accuracy of model predictions, and model hyperparameters. The content regarding the VPT metric, the effects of network thinning on reservoir computing, and the results from grid search experiments mentioned in this thesis has been done previously. The results regarding the diversity metric, kernel tests, and results from BHO are new and use this previous work as a comparison to the quality and usefulness of these new methods in creating accurate reservoir computers.

In this thesis we forecast the future signal strength of base stations in mobile networks. Better forecasts might improve handover of mobile phones between base stations, thus improving overall user experience. Future values are forecast using a series of past sig- nal strength measurements. We use vector autoregression (VAR), a multilayer perceptron (MLP), and a gated recurrent unit (GRU) network. Hyperparameters, including the set of lags, of these models are optimised using Bayesian optimisation (BO) with Gaussian pro- cess (GP) priors. In addition to BO of the VAR model, we optimise the set of lags in it using a standard bottom-up and top-down heuristic. Both approaches result in similar predictive mean squared error (MSE) for the VAR model, but BO requires fewer model estimations. The GRU model provides the best predictive performance out of the three models. How- ever, none of the models (VAR, MLP, or GRU) achieves the accuracy required for practical applicability of the results. Therefore, we suggest adding more information to the model or reformulating the problem.

In the last few years, as processing the data became a part of everyday life in different areas of human activity, the automated machine learning systems that are designed to help with the process of data mining, are on the rise. Various metalearning techniques, including recommendation of the right method to use, or the sequence of steps to take, and to find its optimum hyperparameters configuration, are integrated into these systems to help the researchers with the machine learning tasks. In this thesis, we proposed metalearning algorithms and techniques for hyperparameters optimization, narrowing the intervals of hyperparameters, and recommendations of a machine learning method for a never before seen dataset. We designed two AutoML machine learning systems, where these metalearning techniques are implemented. The extensive set of experiments was proposed to evaluate these algorithms, and the results are presented.