[pt] As técnicas de aprendizado profundo têm mostrado contribuições significativas em vários campos, incluindo a análise de imagens. A grande maioria dos trabalhos em visão computacional concentra-se em propor e aplicar novos modelos e algoritmos de aprendizado de máquina. Para tarefas de aprendizado supervisionado, o desempenho dessas técnicas depende de uma grande quantidade de dados de treinamento, bem como de dados rotulados. No entanto, a rotulagem é um processo caro e demorado. Uma recente área de exploração são as reduções dos esforços na preparação de dados, deixando-os sem inconsistências, ruídos, para que os modelos atuais possam obter um maior desempenho. Esse novo campo de estudo é chamado de Data-Centric IA. Apresentamos uma nova abordagem baseada em Deep Reinforcement Learning (DRL), cujo trabalho é voltado para a preparação de um conjunto de dados em problemas de detecção de objetos, onde as anotações de caixas delimitadoras são feitas de modo autônomo e econômico. Nossa abordagem consiste na criação de uma metodologia para treinamento de um agente virtual a fim de rotular automaticamente os dados, a partir do auxílio humano como professor desse agente. Implementamos o algoritmo Deep Q-Network para criar o agente virtual e desenvolvemos uma abordagem de aconselhamento para facilitar a comunicação do humano professor com o agente virtual estudante. Para completar nossa implementação, utilizamos o método de aprendizado ativo para selecionar casos onde o agente possui uma maior incerteza, necessitando da intervenção humana no processo de anotação durante o treinamento. Nossa abordagem foi avaliada e comparada com outros métodos de aprendizado por reforço e interação humano-computador, em diversos conjuntos de dados, onde o agente virtual precisou criar novas anotações na forma de caixas delimitadoras. Os resultados mostram que o emprego da nossa metodologia impacta positivamente para obtenção de novas anotações a partir de um conjunto de dados com rótulos escassos, superando métodos existentes. Desse modo, apresentamos a contribuição no campo de Data-Centric IA, com o desenvolvimento de uma metodologia de ensino para criação de uma abordagem autônoma com aconselhamento humano para criar anotações econômicas a partir de anotações escassas. / [en] Deep learning techniques have shown significant contributions in various fields, including image analysis. The vast majority of work in computer vision focuses on proposing and applying new machine learning models and algorithms. For supervised learning tasks, the performance of these techniques depends on a large amount of training data and labeled data. However, labeling is an expensive and time-consuming process. A recent area of exploration is the reduction of efforts in data preparation, leaving it without inconsistencies and noise so that current models can obtain greater performance. This new field of study is called Data-Centric AI. We present a new approach based on Deep Reinforcement Learning (DRL), whose work is focused on preparing a dataset, in object detection problems where the bounding box annotations are done autonomously and economically. Our approach consists of creating a methodology for training a virtual agent in order to automatically label the data, using human assistance as a teacher of this agent. We implemented the Deep Q-Network algorithm to create the virtual agent and developed a counseling approach to facilitate the communication of the human teacher with the virtual agent student. We used the active learning method to select cases where the agent has more significant uncertainty, requiring human intervention in the annotation process during training to complete our implementation. Our approach was evaluated and compared with other reinforcement learning methods and human-computer interaction in different datasets, where the virtual agent had to create new annotations in the form of bounding boxes. The results show that the use of our methodology has a positive impact on obtaining new annotations from a dataset with scarce labels, surpassing existing methods. In this way, we present the contribution in the field of Data-Centric AI, with the development of a teaching methodology to create an autonomous approach with human advice to create economic annotations from scarce annotations.

Microprocessor-controlled knees (MPK’s) allow amputees to walk with increasing ease and safety as technology progresses. As an amputee is fitted with a new MPK, the knee’s internal parameters are tuned to the user’s preferred settings in a controlled environment. These parameters determine various gait control settings, such as flexion target angle or swing extension resistance. Though these parameters may work well during the initial fitting, the MPK experiences various internal & external environmental changes throughout its life-cycle, such as product wear, changes in the amputee’s muscle strength, temperature changes, etc. This work investigates the feasibility of using a reinforcement learning (RL) control to adapt the MPK’s swing resistance to consistently induce the amputee’s preferred swing performance in realtime. Three gait features were identified as swing performance indicators for the RL algorithm. Results show that the RL control is able to learn and improve its tuning performance in terms of Mean Absolute Error over two 40-45 minute training sessions with a human-in-the-loop. Additionally, results show promise in using transfer learning to reduce strenuous RL training times. / Mikroprocessorkontrollerade knän (MPK) gör att amputerade kan utföra fysiska aktiviteter med ökad lätthet och säkerhet allt eftersom tekniken fortskrider. När en ny MPK monteras på en amputerad person, anpassas knäts interna parametrar till användarens i ett kontrollerad miljö. Dessa parametrar styr olika gångkontrollinställningar, såsom flexionsmålvinkel eller svängförlängningsmotstånd. Även om parametrarna kan fungera bra under den initiala anpassningen, upplever den MPK olika interna och yttre miljöförändringar under sin hela livscykel, till exempel produktslitage, förändringar i den amputerades muskelstyrka, temperaturförändringar, etc. Detta arbete undersöker möjligheten av, med hjälp av en förstärkningsinlärningskontroll (RL), att anpassa MPK svängmotstånd för att konsekvent inducera den amputerades föredragna svängprestanda i realtid. Tre gångegenskaper identifierades som svingprestandaindikatorer för RL-algoritmen. Resultaten visar att RL-kontrollen kan lära sig och förbättra sin inställningsprestanda i termer av Mean Absolute Error under två 40-45 minuters träningspass med en människa-i-loopen. Dessutom är resultaten lovande när det gäller att använda överföringsinlärning för att minska ansträngande RL-träningstider.

In recent years, Unmanned aerial vehicles (UAVs) have developed rapidly due to advances in aerospace technology, and wireless communication systems. As a result of their versatility, cost-effectiveness, and flexibility of deployment, UAVs have been developed to accomplish a variety of large and complex tasks without terrain restrictions, such as battlefield operations, search and rescue under disaster conditions, monitoring, etc. Data collection and offloading missions in The internet of thingss (IoTs) networks can be accomplished with the use of UAVs as network edge nodes. The fundamental challenge in such scenarios is to develop a UAV movement policy that enhances the quality of mission completion and avoids collisions. Real-time learning based on neural networks has been proven to be an effective method for solving decision-making problems in a dynamic, unknown environment. In this thesis, we assume a real-life scenario in which a UAV collects data from Ground base stations (GBSs) without knowing the information of the environment. A UAV is responsible for the MOO including collecting data, avoiding obstacles, path planning, and conserving energy. Two Deep reinforcement learnings (DRLs) approaches were implemented in this thesis and compared. / Under de senaste åren har UAV utvecklats snabbt på grund av framsteg inom flygteknik och trådlösa kommunikationssystem. Som ett resultat av deras mångsidighet, kostnadseffektivitet och flexibilitet i utbyggnaden har UAV:er utvecklats för att utföra en mängd stora och komplexa uppgifter utan terrängrestriktioner, såsom slagfältsoperationer, sök och räddning under katastrofförhållanden, övervakning, etc. Data insamlings- och avlastningsuppdrag i IoT-nätverk kan utföras med användning av UAV:er som nätverkskantnoder. Den grundläggande utmaningen i sådana scenarier är att utveckla en UAV-rörelsepolicy som förbättrar kvaliteten på uppdragets slutförande och undviker kollisioner. Realtidsinlärning baserad på neurala nätverk har visat sig vara en effektiv metod för att lösa beslutsfattande problem i en dynamisk, okänd miljö. I den här avhandlingen utgår vi från ett verkligt scenario där en UAV samlar in data från GBS utan att känna till informationen om miljön. En UAV är ansvarig för MOO inklusive insamling av data, undvikande av hinder, vägplanering och energibesparing. Två DRL-metoder implementerades i denna avhandling och jämfördes.

L'apprentissage profond par renforcement a produit des décideurs qui jouent aux échecs, au Go, au Shogi, à Atari et à Starcraft avec une capacité surhumaine. Cependant, ces algorithmes ont du mal à naviguer et à contrôler des environnements physiques, contrairement aux animaux et aux humains. Manipuler le monde physique nécessite la maîtrise de domaines d'actions continues tels que la position, la vitesse et l'accélération, contrairement aux domaines d'actions discretes dans des jeux de société et de vidéo. L'entraînement de réseaux neuronaux profonds pour le contrôle continu est instable: les agents ont du mal à apprendre et à conserver de bonnes habitudes, le succès est à haute variance sur hyperparamètres, graines aléatoires, même pour la même tâche, et les algorithmes ont du mal à bien se comporter en dehors des domaines dans lesquels ils ont été développés. Cette thèse examine et améliore l'utilisation de réseaux de neurones profonds dans l'apprentissage par renforcement. Le chapitre 1 explique comment le principe d'entropie maximale produit des fonctions d'objectifs pour l'apprentissage supervisé et non supervisé et déduit, à partir de la dynamique d'apprentissage des réseaux neuronaux profonds, certains termes régulisants pour stabiliser les réseaux neuronaux profonds. Le chapitre 2 fournit une justification de l'entropie maximale pour la forme des algorithmes acteur-critique et trouve une configuration d'un algorithme acteur-critique qui s'entraîne le plus stablement. Enfin, le chapitre 3 examine la dynamique d'apprentissage de l'apprentissage par renforcement profond afin de proposer deux améliorations aux réseaux cibles et jumeaux qui améliorent la stabilité et la convergence. Des expériences sont réalisées dans les simulateurs de physique idéale DeepMind Control, MuJoCo et Box2D. / Deep Reinforcement Learning has produced decision makers that play Chess, Go, Shogi, Atari, and Starcraft with superhuman ability. However, unlike animals and humans, these algorithms struggle to navigate and control physical environments. Manipulating the physical world requires controlling continuous action spaces such as position, velocity, and acceleration, unlike the discrete action spaces of board and video games. Training deep neural networks for continuous control is unstable: agents struggle to learn and retain good behaviors, performance is high variance across hyperparameters, random seed, and even multiple runs of the same task, and algorithms struggle to perform well outside the domains they have been developed in. This thesis finds principles behind the success of deep neural networks in other learning paradigms and examines their impact on reinforcement learning for continuous control. Chapter 1 explains how the maximum-entropy principle produces supervised and unsupervised learning loss functions and derives some regularizers used to stabilize deep networks from the training dynamics of deep learning. Chapter 2 provides a maximum-entropy justification for the form of actor-critic algorithms and finds a configuration of an actor-critic algorithm that trains most stably. Finally, Chapter 3 considers the training dynamics of deep reinforcement learning to propose two improvements to target and twin networks that improve stability and convergence. Experiments are performed within the DeepMind Control, MuJoCo, and Box2D ideal-physics simulators.

L'immense potentiel des approches d'apprentissage par renforcement profond (ARP) pour la conception d'agents autonomes a été démontré à plusieurs reprises au cours de la dernière décennie. Son application à des agents physiques, tels que des robots ou des réseaux électriques automatisés, est cependant confrontée à plusieurs défis. Parmi eux, l'inefficacité de leur échantillonnage, combinée au coût et au risque d'acquérir de l'expérience dans le monde réel, peut décourager tout projet d'entraînement d'agents incarnés. Dans cette thèse, je me concentre sur l'application de l'ARP sur des agents physiques. Je propose d'abord un cadre probabiliste pour améliorer l'efficacité de l'échantillonnage dans l'ARP. Dans un premier article, je présente la pondération BIV (batch inverse-variance), une fonction de perte tenant compte de la variance du bruit des étiquettes dans la régression bruitée hétéroscédastique. La pondération BIV est un élément clé du deuxième article, où elle est combinée avec des méthodes de pointe de prédiction de l'incertitude pour les réseaux neuronaux profonds dans un pipeline bayésien pour les algorithmes d'ARP avec différences temporelles. Cette approche, nommée apprentissage par renforcement à variance inverse (IV-RL), conduit à un entraînement nettement plus rapide ainsi qu'à de meilleures performances dans les tâches de contrôle. Dans le troisième article, l'apprentissage par renforcement multi-agent (MARL) est appliqué au problème de la réponse rapide à la demande, une approche prometteuse pour gérer l'introduction de sources d'énergie renouvelables intermittentes dans les réseaux électriques. En contrôlant la coordination de plusieurs climatiseurs, les agents MARL obtiennent des performances nettement supérieures à celles des approches basées sur des règles. Ces résultats soulignent le rôle potentiel que les agents physiques entraînés par MARL pourraient jouer dans la transition énergétique et la lutte contre le réchauffement climatique. / The immense potential of deep reinforcement learning (DRL) approaches to build autonomous agents has been proven repeatedly in the last decade. Its application to embodied agents, such as robots or automated power systems, is however facing several challenges. Among them, their sample inefficiency, combined to the cost and the risk of gathering experience in the real world, can deter any idea of training embodied agents. In this thesis, I focus on the application of DRL on embodied agents. I first propose a probabilistic framework to improve sample efficiency in DRL. In the first article, I present batch inverse-variance (BIV) weighting, a loss function accounting for label noise variance in heteroscedastic noisy regression. BIV is a key element of the second article, where it is combined with state-of-the-art uncertainty prediction methods for deep neural networks in a Bayesian pipeline for temporal differences DRL algorithms. This approach, named inverse-variance reinforcement learning (IV-RL), leads to significantly faster training as well as better performance in control tasks. In the third article, multi-agent reinforcement learning (MARL) is applied to the problem of fast-timescale demand response, a promising approach to the manage the introduction of intermittent renewable energy sources in power-grids. As MARL agents control the coordination of multiple air conditioners, they achieve significantly better performance than rule-based approaches. These results underline to the potential role that DRL trained embodied agents could take in the energetic transition and the fight against global warming.

[pt] Em ambientes aquáticos, o uso tradicional de mergulhadores ou veiculos subaquáticos tripulados foi substituído por veículos subaquáticos não tripulados (como ROVs ou AUVs). Com vantagens em termos de redução de riscos de segurança, como exposição à pressão, temperatura ou falta de ar. Além disso, conseguem acessar áreas de extrema profundidade que até então não eram possiveis para o ser humano. Esses veiculos não tripulados são amplamente utilizados para inspeções como as necessárias para o descomissionamento de plataformas de petróleo Neste tipo de fiscalização é necessário analisar as condições do solo, da tu- bulação e, principalmente, se foi criado um ecossistema próximo à tubulação. Grande parte dos trabalhos realizados para a automação desses veículos utilizam diferentes tipos de sensores e GPS para realizar a percepção do ambiente. Devido à complexidade do ambiente de navegação, diferentes algoritmos de controle e automação têm sido testados nesta área, O interesse deste trabalho é fazer com que o autômato tome decisões através da análise de eventos visuais. Este método de pesquisa traz a vantagem de redução de custos para o projeto, visto que as câmeras possuem um preço inferior em relação aos sensores ou dispositivos GPS. A tarefa de inspeção autônoma tem vários desafios: detectar os eventos, processar as imagens e tomar a decisão de alterar a rota em tempo real. É uma tarefa altamente complexa e precisa de vários algoritmos trabalhando juntos para ter um bom desempenho. A inteligência artificial apresenta diversos algoritmos para automatizar, como os baseados em aprendizagem por reforço entre outros na área de detecção e classificação de imagens Esta tese de doutorado consiste em um estudo para criação de um sistema avançado de inspeção autônoma. Este sistema é capaz de realizar inspeções apenas analisando imagens da câmera AUV, usando aprendizagem de reforço profundo profundo para otimizar o planejamento do ponto de vista e técnicas de detecção de novidades. Contudo, este quadro pode ser adaptado a muitas outras tarefas de inspecção. Neste estudo foram utilizados ambientes realistas complexos, nos quais o agente tem o desafio de chegar da melhor forma possível ao objeto de interesse para que possa classificar o objeto. Vale ressaltar, entretanto, que os ambientes de simulação utilizados neste contexto apresentam certo grau de simplicidade carecendo de recursos como correntes marítimas on dinâmica de colisão em seus cenários simulados Ao final deste projeto, o Visual Inspection of Pipelines (VIP) framework foi desenvolvido e testado, apresentando excelentes resultados e ilustrando a viabilidade de redução do tempo de inspeção através da otimização do planejamento do ponto de vista. Esse tipo de abordagem, além de agregar conhecimento ao robô autônomo, faz com que as inspeções subaquáticas exijam pouca presença de ser humano (human-in-the-loop), justificando o uso das técnicas empregadas. / [en] In aquatic environments, the traditional use of divers or manned underwater vehicles has been replaced by unmanned underwater vehicles (such as ROVs or AUVs). With advantages in terms of reducing safety risks, such as exposure to pressure, temperature or shortness of breath. In addition, they are able to access areas of extreme depth that were not possible for humans until then. These unmanned vehicles are widely used for inspections, such as those required for the decommissioning of oil platforms. In this type of inspection, it is necessary to analyze the conditions of the soil, the pipeline and, especially, if an ecosystem was created close to the pipeline. Most of the works carried out for the automation of these vehicles use different types of sensors and GPS to perform the perception of the environment. Due to the complexity of the navigation environment, different control and automation algorithms have been tested in this area. The interest of this work is to make the automaton take decisions through the analysis of visual events. This research method provides the advantage of cost reduction for the project, given that cameras have a lower price compared to sensors or GPS devices. The autonomous inspection task has several challenges: detecting the events, processing the images and making the decision to change the route in real time. It is a highly complex task and needs multiple algorithms working together to perform well. Artificial intelligence presents many algorithms to automate, such as those based on reinforcement learning, among others in the area of image detection and classification. This doctoral thesis consists of a study to create an advanced autonomous inspection system. This system is capable of performing inspections only by analyzing images from the AUV camera, using deep reinforcement learning, and novelty detection techniques. However, this framework can be adapted to many other inspection tasks. In this study, complex realistic environments were used, in which the agent has the challenge of reaching the object of interest in the best possible way so that it can classify the object. It is noteworthy, however, that the simulation environments utilized in this context exhibit a certain degree of simplicity, lacking features like marine currents or collision dynamics in their simulated scenarios. At the conclusion of this project, a Visual Inspection of Pipelines (VIP) framework was developed and tested, showcasing excellent results and illustrating the feasibility of reducing inspection time through the optimization of viewpoint planning. This type of approach, in addition to adding knowledge to the autonomous robot, means that underwater inspections require little pres- ence of a human being (human-in-the-loop), justifying the use of the techniques employed.

In this project we investigate the susceptibility ofreinforcement rearning (RL) algorithms to adversarial attacks.Adversarial attacks have been proven to be very effective atreducing performance of deep learning classifiers, and recently,have also been shown to reduce performance of RL agents.The goal of this project is to evaluate adversarial attacks onagents trained using deep reinforcement learning (DRL), aswell as to investigate how to detect these types of attacks. Wefirst use DRL to solve two environments from OpenAI’s gymmodule, namely Cartpole and Lunarlander, by using DQN andDDPG (DRL techniques). We then evaluate the performanceof attacks and finally we also train neural networks to detectattacks. The attacks was successful at reducing performancein the LunarLander environment and CartPole environment.The attack detector was very successful at detecting attacks onthe CartPole environment, but performed not quiet as well onLunarLander.We hypothesize that continuous action space environmentsmay pose a greater difficulty for attack detectors to identifypotential adversarial attacks. / I detta projekt undersöker vikänsligheten hos förstärknings lärda (RL) algotritmerför attacker mot förstärknings lärda agenter. Attackermot förstärknings lärda agenter har visat sig varamycket effektiva för att minska prestandan hos djuptförsärknings lärda klassifierare och har nyligen visat sigockså minska prestandan hos förstärknings lärda agenter.Målet med detta projekt är att utvärdera attacker motdjupt förstärknings lärda agenter och försöka utföraoch upptäcka attacker. Vi använder först RL för attlösa två miljöer från OpenAIs gym module CartPole-v0och ContiniousLunarLander-v0 med DQN och DDPG.Vi utvärderar sedan utförandet av attacker och avslutarslutligen med ett möjligt sätt att upptäcka attacker.Attackerna var mycket framgångsrika i att minskaprestandan i både CartPole-miljön och LunarLandermiljön. Attackdetektorn var mycket framgångsrik medatt upptäcka attacker i CartPole-miljön men presteradeinte lika bra i LunarLander-miljön.Vi hypotiserar att miljöer med kontinuerligahandlingsrum kan innebära en större svårighet fören attack identifierare att upptäcka attacker mot djuptförstärknings lärda agenter. / Kandidatexjobb i elektroteknik 2021, KTH, Stockholm

<p dir="ltr">Current artificial intelligence (AI) systems are proficient at tasks in a closed-world setting where the rules are often rigid. However, in real-world applications, the environment is usually open and dynamic. In this work, we investigate the effects of such dynamic environments on AI systems and develop ways to mitigate those effects. Central to our exploration is the concept of \textit{novelties}. Novelties encompass structural changes, unanticipated events, and environmental shifts that can confound traditional AI systems. We categorize novelties based on their representation, anticipation, and impact on agents, laying the groundwork for systematic detection and adaptation strategies. We explore novelties in the context of stochastic games. Decision-making in stochastic games exercises many aspects of the same reasoning capabilities needed by AI agents acting in the real world. A multi-agent stochastic game allows for infinitely many ways to introduce novelty. We propose an extension of the deep reinforcement learning (DRL) paradigm to develop agents that can detect and adapt to novelties in these environments. To address the sample efficiency challenge in DRL, we introduce a hybrid approach that combines fixed-policy methods with traditional DRL techniques, offering enhanced performance in complex decision-making tasks. We present a novel method for detecting anticipated novelties in multi-agent games, leveraging information theory to discern patterns indicative of collusion among players. Finally, we introduce DABLER, a pioneering deep reinforcement learning architecture that dynamically adapts to changing environmental conditions through broad learning approaches and environment recognition. Our findings underscore the importance of developing AI systems equipped to navigate the uncertainties of the open world, offering promising pathways for advancing AI research and application in real-world settings.</p>

L’apprentissage par renforcement profond (RL) a connu d’énormes progrès ces dernières années, mais il est encore difficile d’appliquer le RL aux problèmes de prise de décision du monde réel. Cette thèse identifie trois défis clés avec la façon dont nous faisons la recherche RL elle-même qui entravent les progrès de la recherche RL. — Évaluation et comparaison peu fiables des algorithmes RL ; les méthodes d’évaluation actuelles conduisent souvent à des résultats peu fiables. — Manque d’informations préalables dans la recherche RL ; Les algorithmes RL sont souvent formés à partir de zéro, ce qui peut nécessiter de grandes quantités de données ou de ressources informatiques. — Manque de compréhension de la façon dont les réseaux de neurones profonds interagissent avec RL, ce qui rend difficile le développement de méthodes évolutives de RL. Pour relever ces défis susmentionnés, cette thèse apporte les contributions suivantes : — Une méthodologie plus rigoureuse pour évaluer les algorithmes RL. — Un flux de travail de recherche alternatif qui se concentre sur la réutilisation des progrès existants sur une tâche. — Identification d’un phénomène de perte de capacité implicite avec un entraînement RL hors ligne prolongé. Dans l’ensemble, cette thèse remet en question le statu quo dans le RL profond et montre comment cela peut conduire à des algorithmes de RL plus efficaces, fiables et mieux applicables dans le monde réel. / Deep reinforcement learning (RL) has seen tremendous progress in recent years, but it is still difficult to apply RL to real-world decision-making problems. This thesis identifies three key challenges with how we do RL research itself that hinder the progress of RL research. — Unreliable evaluation and comparison of RL algorithms; current evaluation methods often lead to unreliable results. — Lack of prior information in RL research; RL algorithms are often trained from scratch, which can require large amounts of data or computational resources. — Lack of understanding of how deep neural networks interact with RL, making it hard to develop scalable RL methods. To tackle these aforementioned challenges, this thesis makes the following contributions: — A more rigorous methodology for evaluating RL algorithms. — An alternative research workflow that focuses on reusing existing progress on a task. — Identifying an implicit capacity loss phenomenon with prolonged offline RL training. Overall, this thesis challenges the status quo in deep reinforcement learning and shows that doing so can make RL more efficient, reliable and improve its real-world applicability

The work in this thesis is separated into two parts. The first part deals with the load planning and sequencing problem for double-stack intermodal railcars, an operational problem found at many rail container terminals. In this problem, containers must be assigned to a platform on which the container will be loaded, and the loading order must be determined. These decisions are made with the objective of minimizing the costs associated with handling the containers, as well as minimizing the cost of containers left behind. The deterministic version of the problem can be cast as a shortest path problem on an ordered graph. This problem is challenging to solve because of the large size of the graph. We propose a two-stage heuristic based on the Iterative Deepening A* algorithm to compute solutions to the load planning and sequencing problem within a five-minute time budget. Next, we also illustrate how a Deep Q-learning algorithm can be used to heuristically solve the same problem.The second part of this thesis considers sequential models in deep learning. A recent strategy to circumvent the exploding and vanishing gradient problem in recurrent neural networks (RNNs) is to enforce recurrent weight matrices to be orthogonal or unitary. While this ensures stable dynamics during training, it comes at the cost of reduced expressivity due to the limited variety of orthogonal transformations. We propose a parameterization of RNNs, based on the Schur decomposition, that mitigates the exploding and vanishing gradient problem, while allowing for non-orthogonal recurrent weight matrices in the model. / Le travail de cette thèse est divisé en deux parties. La première partie traite du problème de planification et de séquencement des chargements de conteneurs sur des wagons, un problème opérationnel rencontré dans de nombreux terminaux ferroviaires intermodaux. Dans ce problème, les conteneurs doivent être affectés à une plate-forme sur laquelle un ou deux conteneurs seront chargés et l'ordre de chargement doit être déterminé. Ces décisions sont prises dans le but de minimiser les coûts associés à la manutention des conteneurs, ainsi que de minimiser le coût des conteneurs non chargés. La version déterministe du problème peut être formulé comme un problème de plus court chemin sur un graphe ordonné. Ce problème est difficile à résoudre en raison de la grande taille du graphe. Nous proposons une heuristique en deux étapes basée sur l'algorithme Iterative Deepening A* pour calculer des solutions au problème de planification et de séquencement de la charge dans un budget de cinq minutes. Ensuite, nous illustrons également comment un algorithme d'apprentissage Deep Q peut être utilisé pour résoudre heuristiquement le même problème. La deuxième partie de cette thèse examine les modèles séquentiels en apprentissage profond. Une stratégie récente pour contourner le problème de gradient qui explose et disparaît dans les réseaux de neurones récurrents (RNN) consiste à imposer des matrices de poids récurrentes orthogonales ou unitaires. Bien que cela assure une dynamique stable pendant l'entraînement, cela se fait au prix d'une expressivité réduite en raison de la variété limitée des transformations orthogonales. Nous proposons une paramétrisation des RNN, basée sur la décomposition de Schur, qui atténue les problèmes de gradient, tout en permettant des matrices de poids récurrentes non orthogonales dans le modèle.