Problems involving optimal sequential making in uncertain dynamic systems arise in domains such as engineering, science and economics. Such problems can often be cast in the framework of Markov Decision Process (MDP). Solving an MDP requires computing the optimal value function and the optimal policy. The idea of dynamic programming (DP) and the Bellman equation (BE) are at the heart of solution methods. The three important exact DP methods are value iteration, policy iteration and linear programming. The exact DP methods compute the optimal value function and the optimal policy. However, the exact DP methods are inadequate in practice because the state space is often large and in practice, one might have to resort to approximate methods that compute sub-optimal policies. Further, in certain cases, the system observations are known only in the form of noisy samples and we need to design algorithms that learn from these samples. In this thesis we study interesting theoretical questions pertaining to approximate and learning algorithms, and also present an interesting application of MDPs in the domain of crowd sourcing. Approximate Dynamic Programming (ADP) methods handle the issue of large state space by computing an approximate value function and/or a sub-optimal policy. In this thesis, we are concerned with conditions that result in provably good policies. Motivated by the limitations of the PBE in the conventional linear algebra, we study the PBE in the (min, +) linear algebra. It is a well known fact that deterministic optimal control problems with cost/reward criterion are (min, +)/(max, +) linear and ADP methods have been developed for such systems in literature. However, it is straightforward to show that infinite horizon discounted reward/cost MDPs are neither (min, +) nor (max, +) linear. We develop novel ADP schemes namely the Approximate Q Iteration (AQI) and Variational Approximate Q Iteration (VAQI), where the approximate solution is a (min, +) linear combination of a set of basis functions whose span constitutes a subsemimodule. We show that the new ADP methods are convergent and we present a bound on the performance of the sub-optimal policy. The Approximate Linear Program (ALP) makes use of linear function approximation (LFA) and offers theoretical performance guarantees. Nevertheless, the ALP is difficult to solve due to the presence of a large number of constraints and in practice, a reduced linear program (RLP) is solved instead. The RLP has a tractable number of constraints sampled from the original constraints of the ALP. Though the RLP is known to perform well in experiments, theoretical guarantees are available only for a specific RLP obtained under idealized assumptions. In this thesis, we generalize the RLP to define a generalized reduced linear program (GRLP) which has a tractable number of constraints that are obtained as positive linear combinations of the original constraints of the ALP. The main contribution here is the novel theoretical framework developed to obtain error bounds for any given GRLP. Reinforcement Learning (RL) algorithms can be viewed as sample trajectory based solution methods for solving MDPs. Typically, RL algorithms that make use of stochastic approximation (SA) are iterative schemes taking small steps towards the desired value at each iteration. Actor-Critic algorithms form an important sub-class of RL algorithms, wherein, the critic is responsible for policy evaluation and the actor is responsible for policy improvement. The actor and critic iterations have deferent step-size schedules, in particular, the step-sizes used by the actor updates have to be generally much smaller than those used by the critic updates. Such SA schemes that use deferent step-size schedules for deferent sets of iterates are known as multitimescale stochastic approximation schemes. One of the most important conditions required to ensure the convergence of the iterates of a multi-timescale SA scheme is that the iterates need to be stable, i.e., they should be uniformly bounded almost surely. However, the conditions that imply the stability of the iterates in a multi-timescale SA scheme have not been well established. In this thesis, we provide veritable conditions that imply stability of two timescale stochastic approximation schemes. As an example, we also demonstrate that the stability of a widely used actor-critic RL algorithm follows from our analysis. Crowd sourcing (crowd) is a new mode of organizing work in multiple groups of smaller chunks of tasks and outsourcing them to a distributed and large group of people in the form of an open call. Recently, crowd sourcing has become a major pool for human intelligence tasks (HITs) such as image labeling, form digitization, natural language processing, machine translation evaluation and user surveys. Large organizations/requesters are increasingly interested in crowd sourcing the HITs generated out of their internal requirements. Task starvation leads to huge variation in the completion times of the tasks posted on to the crowd. This is an issue for frequent requesters desiring predictability in the completion times of tasks specified in terms of percentage of tasks completed within a stipulated amount of time. An important task attribute that affects the completion time of a task is its price. However, a pricing policy that does not take the dynamics of the crowd into account might fail to achieve the desired predictability in completion times. Here, we make use of the MDP framework to compute a pricing policy that achieves predictable completion times in simulations as well as real world experiments.

Cette thèse porte sur la résolution de problèmes de décision séquentielle sous incertitude,modélisés sous forme de processus décisionnels de Markov (PDM) dont l’espace d’étatet d’action sont tous les deux de grande dimension. La résolution de ces problèmes avecun bon compromis entre qualité de l’approximation et passage à l’échelle est encore unchallenge. Les algorithmes de résolution dédiés à ce type de problèmes sont rares quandla dimension des deux espaces excède 30, et imposent certaines limites sur la nature desproblèmes représentables.Nous avons proposé un nouveau cadre, appelé PDMF3, ainsi que des algorithmesde résolution approchée associés. Un PDMF3 est un processus décisionnel de Markov àespace d’état et d’action factorisés (PDMF-AF) dont non seulement l’espace d’état etd’action sont factorisés mais aussi dont les politiques solutions sont contraintes à unecertaine forme factorisée, et peuvent être stochastiques. Les algorithmes que nous avonsproposés appartiennent à la famille des algorithmes de type itération de la politique etexploitent des techniques d’optimisation continue et des méthodes d’inférence dans lesmodèles graphiques. Ces algorithmes de type itération de la politique ont été validés sur un grand nombre d’expériences numériques. Pour de petits PDMF3, pour lesquels la politique globale optimale est disponible, ils fournissent des politiques solutions proches de la politique globale optimale. Pour des problèmes plus grands de la sous-classe des processus décisionnels de Markov sur graphe (PDMG), ils sont compétitifs avec des algorithmes de résolution de l’état de l’art en termes de qualité. Nous montrons aussi que nos algorithmes permettent de traiter des PDMF3 de très grande taille en dehors de la sous-classe des PDMG, sur des problèmes jouets inspirés de problèmes réels en agronomie ou écologie. L’espace d’état et d’action sont alors tous les deux de dimension 100, et de taille 2100. Dans ce cas, nous comparons la qualité des politiques retournées à celle de politiques expertes. Dans la seconde partie de la thèse, nous avons appliqué le cadre et les algorithmesproposés pour déterminer des stratégies de gestion des services écosystémiques dans unpaysage agricole. Les adventices, plantes sauvages des milieux agricoles, présentent desfonctions antagonistes, étant à la fois en compétition pour les ressources avec la cultureet à la base de réseaux trophiques dans les agroécosystèmes. Nous cherchons à explorerquelles organisations du paysage (ici composé de colza, blé et prairie) dans l’espace etdans le temps permettent de fournir en même temps des services de production (rendementen céréales, fourrage et miel), des services de régulation (régulation des populationsd’espèces adventices et de pollinisateurs sauvages) et des services culturels (conservationd’espèces adventices et de pollinisateurs sauvages). Pour cela, nous avons développé unmodèle de la dynamique des adventices et des pollinisateurs et de la fonction de récompense pour différents objectifs (production, maintien de la biodiversité ou compromisentre les services). L’espace d’état de ce PDMF3 est de taille 32100, et l’espace d’actionde taille 3100, ce qui en fait un problème de taille conséquente. La résolution de ce PDMF3 a conduit à identifier différentes organisations du paysage permettant d’atteindre différents bouquets de services écosystémiques, qui diffèrent dans la magnitude de chacune des trois classes de services écosystémiques. / This PhD thesis focuses on the resolution of problems of sequential decision makingunder uncertainty, modelled as Markov decision processes (MDP) whose state and actionspaces are both of high dimension. Resolution of these problems with a good compromisebetween quality of approximation and scaling is still a challenge. Algorithms for solvingthis type of problems are rare when the dimension of both spaces exceed 30, and imposecertain limits on the nature of the problems that can be represented.We proposed a new framework, called F3MDP, as well as associated approximateresolution algorithms. A F3MDP is a Markov decision process with factored state andaction spaces (FA-FMDP) whose solution policies are constrained to be in a certainfactored form, and can be stochastic. The algorithms we proposed belong to the familyof approximate policy iteration algorithms and make use of continuous optimisationtechniques, and inference methods for graphical models.These policy iteration algorithms have been validated on a large number of numericalexperiments. For small F3MDPs, for which the optimal global policy is available, theyprovide policy solutions that are close to the optimal global policy. For larger problemsfrom the graph-based Markov decision processes (GMDP) subclass, they are competitivewith state-of-the-art algorithms in terms of quality. We also show that our algorithmsallow to deal with F3MDPs of very large size outside the GMDP subclass, on toy problemsinspired by real problems in agronomy or ecology. The state and action spaces arethen both of dimension 100, and of size 2100. In this case, we compare the quality of thereturned policies with the one of expert policies. In the second part of the thesis, we applied the framework and the proposed algorithms to determine ecosystem services management strategies in an agricultural landscape.Weed species, ie wild plants of agricultural environments, have antagonistic functions,being at the same time in competition with the crop for resources and keystonespecies in trophic networks of agroecosystems. We seek to explore which organizationsof the landscape (here composed of oilseed rape, wheat and pasture) in space and timeallow to provide at the same time production services (production of cereals, fodder andhoney), regulation services (regulation of weed populations and wild pollinators) andcultural services (conservation of weed species and wild pollinators). We developed amodel for weeds and pollinators dynamics and for reward functions modelling differentobjectives (production, conservation of biodiversity or trade-off between services). Thestate space of this F3MDP is of size 32100, and the action space of size 3100, which meansthis F3MDP has substantial size. By solving this F3MDP, we identified various landscapeorganizations that allow to provide different sets of ecosystem services which differ inthe magnitude of each of the three classes of ecosystem services.

In many optimization problems, the relationship between the objective and parameters is not known. The objective function itself may be stochastic such as a long-run average over some random cost samples. In such cases finding the gradient of the objective is not possible. It is in this setting that stochastic approximation algorithms are used. These algorithms use some estimates of the gradient and are stochastic in nature. Amongst gradient estimation techniques, Simultaneous Perturbation Stochastic Approximation (SPSA) and Smoothed Functional(SF) scheme are widely used. In this thesis we have proposed a novel multi-time scale quasi-Newton based smoothed functional (QN-SF) algorithm for unconstrained as well as constrained optimization. The algorithm uses the smoothed functional scheme for estimating the gradient and the quasi-Newton method to solve the optimization problem. The algorithm is shown to converge with probability one. We have also provided here experimental results on the problem of optimal routing in a multi-stage network of queues. Policies like Join the Shortest Queue or Least Work Left assume knowledge of the queue length values that can change rapidly or hard to estimate. If the only information available is the expected end-to-end delay as with our case, such policies cannot be used. The QN-SF based probabilistic routing algorithm uses only the total end-to-end delay for tuning the probabilities. We observe from the experiments that the QN-SF algorithm has better performance than the gradient and Jacobi versions of Newton based smoothed functional algorithms. Next we consider constrained routing in a similar queueing network. We extend the QN-SF algorithm to this case. We study the convergence behavior of the algorithm and observe that the constraints are satisfied at the point of convergence. We provide experimental results for the constrained routing setup as well. Next we study reinforcement learning algorithms which are useful for solving Markov Decision Process(MDP) when the precise information on transition probabilities is not known. When the state, and action sets are very large, it is not possible to store all the state-action tuples. In such cases, function approximators like neural networks have been used. The popular Q-learning algorithm is known to diverge when used with linear function approximation due to the ’off-policy’ problem. Hence developing stable learning algorithms when used with function approximation is an important problem. We present in this thesis a variant of Q-learning with linear function approximation that is based on two-timescale stochastic approximation. The Q-value parameters for a given policy in our algorithm are updated on the slower timescale while the policy parameters themselves are updated on the faster scale. We perform a gradient search in the space of policy parameters. Since the objective function and hence the gradient are not analytically known, we employ the efficient one-simulation simultaneous perturbation stochastic approximation(SPSA) gradient estimates that employ Hadamard matrix based deterministic perturbations. Our algorithm has the advantage that, unlike Q-learning, it does not suffer from high oscillations due to the off-policy problem when using function approximators. Whereas it is difficult to prove convergence of regular Q-learning with linear function approximation because of the off-policy problem, we prove that our algorithm which is on-policy is convergent. Numerical results on a multi-stage stochastic shortest path problem show that our algorithm exhibits significantly better performance and is more robust as compared to Q-learning. Future work would be to compare it with other policy-based reinforcement learning algorithms. Finally, we develop an online actor-critic reinforcement learning algorithm with function approximation for a problem of control under inequality constraints. We consider the long-run average cost Markov decision process(MDP) framework in which both the objective and the constraint functions are suitable policy-dependent long-run averages of certain sample path functions. The Lagrange multiplier method is used to handle the inequality constraints. We prove the asymptotic almost sure convergence of our algorithm to a locally optimal solution. We also provide the results of numerical experiments on a problem of routing in a multistage queueing network with constraints on long-run average queue lengths. We observe that our algorithm exhibits good performance on this setting and converges to a feasible point.

Machine learning is spearheading progress for the field of artificial intelligence in terms of providing competition in strategy games to a human opponent, be it in a game of chess, Go or poker. A field of machine learning, which shows the most promising results in playing strategy games, is reinforcement learning. The next milestone for the current research lies in a computer game Starcraft II, which outgrows the previous ones in terms of complexity, and represents a potential new breakthrough in this field. The paper focuses on analysis of the problem, and suggests a solution incorporating a reinforcement learning algorithm A2C and hyperparameter optimization implementation PBT, which could mean a step forward for the current progress.

Une grande partie de ces données volumineuses ont plus de valeur lorsqu'elles sont analysées rapidement, au fur et à mesure de leur génération. Dans plusieurs scénarios d'application émergents, tels que les villes intelligentes, la surveillance opérationnelle de grandes infrastructures et l'Internet des Objets (Internet of Things), des flux continus de données doivent être traités dans des délais très brefs. Dans plusieurs domaines, ce traitement est nécessaire pour détecter des modèles, identifier des défaillances et pour guider la prise de décision. Les données sont donc souvent rassemblées et analysées par des environnements logiciels conçus pour le traitement de flux continus de données. Ces environnements logiciels pour le traitement de flux de données déploient les applications sous-la forme d'un graphe orienté ou de dataflow. Un dataflow contient une ou plusieurs sources (i.e. capteurs, passerelles ou actionneurs); opérateurs qui effectuent des transformations sur les données (e.g., filtrage et agrégation); et des sinks (i.e., éviers qui consomment les requêtes ou stockent les données). Nous proposons dans cette thèse un ensemble de stratégies pour placer les opérateurs dans une infrastructure massivement distribuée cloud-edge en tenant compte des caractéristiques des ressources et des exigences des applications. En particulier, nous décomposons tout d'abord le graphe d'application en identifiant quelques comportements tels que des forks et des joints, puis nous le plaçons dynamiquement sur l'infrastructure. Des simulations et un prototype prenant en compte plusieurs paramètres d'application démontrent que notre approche peut réduire la latence de bout en bout de plus de 50% et aussi améliorer d'autres métriques de qualité de service. L'espace de recherche de solutions pour la reconfiguration des opérateurs peut être énorme en fonction du nombre d'opérateurs, de flux, de ressources et de liens réseau. De plus, il est important de minimiser le coût de la migration tout en améliorant la latence. Des travaux antérieurs, Reinforcement Learning (RL) et Monte-Carlo Tree Searh (MCTS) ont été utilisés pour résoudre les problèmes liés aux grands nombres d’actions et d’états de recherche. Nous modélisons le problème de reconfiguration d'applications sous la forme d'un processus de décision de Markov (MDP) et étudions l'utilisation des algorithmes RL et MCTS pour concevoir des plans de reconfiguration améliorant plusieurs métriques de qualité de service. / A large part of this big data is most valuable when analysed quickly, as it is generated. Under several emerging application scenarios, such as in smart cities, operational monitoring of large infrastructure, and Internet of Things (IoT), continuous data streams must be processed under very short delays. In multiple domains, there is a need for processing data streams to detect patterns, identify failures, and gain insights. Data is often gathered and analysed by Data Stream Processing Engines (DSPEs).A DSPE commonly structures an application as a directed graph or dataflow. A dataflow has one or multiple sources (i.e., gateways or actuators); operators that perform transformations on the data (e.g., filtering); and sinks (i.e., queries that consume or store the data). Most complex operator transformations store information about previously received data as new data is streamed in. Also, a dataflow has stateless operators that consider only the current data. Traditionally, Data Stream Processing (DSP) applications were conceived to run in clusters of homogeneous resources or on the cloud. In a cloud deployment, the whole application is placed on a single cloud provider to benefit from virtually unlimited resources. This approach allows for elastic DSP applications with the ability to allocate additional resources or release idle capacity on demand during runtime to match the application requirements.We introduce a set of strategies to place operators onto cloud and edge while considering characteristics of resources and meeting the requirements of applications. In particular, we first decompose the application graph by identifying behaviours such as forks and joins, and then dynamically split the dataflow graph across edge and cloud. Comprehensive simulations and a real testbed considering multiple application settings demonstrate that our approach can improve the end-to-end latency in over 50% and even other QoS metrics. The solution search space for operator reassignment can be enormous depending on the number of operators, streams, resources and network links. Moreover, it is important to minimise the cost of migration while improving latency. Reinforcement Learning (RL) and Monte-Carlo Tree Search (MCTS) have been used to tackle problems with large search spaces and states, performing at human-level or better in games such as Go. We model the application reconfiguration problem as a Markov Decision Process (MDP) and investigate the use of RL and MCTS algorithms to devise reconfiguring plans that improve QoS metrics.

<p>It is predicted that there would be 41.6 billion IoT devices by 2025, which has kindled new interests on the timing coordination between sensors and controllers, i.e., how to use the waiting time to improve the performance. Sun et al. showed that a <i>controller</i> can strictly improve the data freshness, the so-called Age-of-Information (AoI), via careful scheduling designs. The optimal waiting policy for the <i>sensor</i> side was later characterized in the context of remote estimation. The first part of this work develops the jointly optimal sensor/controller waiting policy. It generalizes the above two important results in that not only do we consider joint sensor/controller designs, but we also assume random delay in both the forward and feedback directions. </p> <p> </p> <p>The second part of the work revisits and significantly strengthens the seminal results of Sun et al on the following fronts: (i) When designing the optimal offline schemes with full knowledge of the delay distributions, a new <i>fixed-point-based</i> method is proposed with <i>quadratic convergence rate</i>; (ii) When the distributional knowledge is unavailable, two new low-complexity online algorithms are proposed, which provably attain the optimal average AoI penalty; and (iii) the online schemes also admit a modular architecture, which allows the designer to <i>upgrade</i> certain components to handle additional practical challenges. Two such upgrades are proposed: (iii.1) the AoI penalty function incurred at the destination is unknown to the source node and must also be estimated on the fly, and (iii.2) the unknown delay distribution is Markovian instead of i.i.d. </p> <p> </p> <p>With the exponential growth of interconnected IoT devices and the increasing risk of excessive resource consumption in mind, the third part of this work derives an optimal joint cost-and-AoI minimization solution for multiple coexisting source-destination (S-D) pairs. The results admit a new <i>AoI-market-price</i>-based interpretation and are applicable to the setting of (i) general heterogeneous AoI penalty functions and Markov delay distributions for each S-D pair, and (ii) a general network cost function of aggregate throughput of all S-D pairs. </p> <p> </p> <p>In each part of this work, extensive simulation is used to demonstrate the superior performance of the proposed schemes. The discussion on analytical as well as numerical results sheds some light on designing practical network utility maximization protocols.</p>

L'apprentissage par renforcement (RL) est un paradigme d'apprentissage orienté agent qui s'intéresse à l'apprentissage en interagissant avec un environnement incertain. Combiné à des réseaux de neurones profonds comme approximateur de fonction, l'apprentissage par renforcement profond (Deep RL) nous a permis récemment de nous attaquer à des tâches très complexes et de permettre à des agents artificiels de maîtriser des jeux classiques comme le Go, de jouer à des jeux vidéo à partir de pixels et de résoudre des tâches de contrôle robotique. Toutefois, un examen plus approfondi de ces remarquables succès empiriques révèle certaines limites fondamentales. Tout d'abord, il a été difficile de combiner les caractéristiques souhaitables des algorithmes RL, telles que l'apprentissage hors politique et en plusieurs étapes, et l'approximation de fonctions, de manière à obtenir des algorithmes stables et efficaces dans de grands espaces d'états. De plus, les algorithmes RL profonds ont tendance à être très inefficaces en raison des stratégies d'exploration-exploitation rudimentaires que ces approches emploient. Enfin, ils nécessitent une énorme quantité de données supervisées et finissent par produire un agent étroit capable de résoudre uniquement la tâche sur laquelle il est entrainé. Dans cette thèse, nous proposons de nouvelles solutions aux problèmes de l'apprentissage hors politique et du dilemme exploration-exploitation dans les grands espaces d'états, ainsi que de l'auto-supervision dans la RL. En ce qui concerne l'apprentissage hors politique, nous apportons deux contributions. Tout d'abord, pour le problème de l'évaluation des politiques, nous montrons que la combinaison des méthodes populaires d'apprentissage hors politique et à plusieurs étapes avec une paramétrisation linéaire de la fonction de valeur pourrait conduire à une instabilité indésirable, et nous dérivons une variante de ces méthodes dont la convergence est prouvée. Deuxièmement, pour l'optimisation des politiques, nous proposons de stabiliser l'étape d'amélioration des politiques par une régularisation de divergence hors politique qui contraint les distributions stationnaires d'états induites par des politiques consécutives à être proches les unes des autres. Ensuite, nous étudions l'apprentissage en ligne dans de grands espaces d'états et nous nous concentrons sur deux hypothèses structurelles pour rendre le problème traitable : les environnements lisses et linéaires. Pour les environnements lisses, nous proposons un algorithme en ligne efficace qui apprend activement un partitionnement adaptatif de l'espace commun en zoomant sur les régions les plus prometteuses et fréquemment visitées. Pour les environnements linéaires, nous étudions un cadre plus réaliste, où l'environnement peut maintenant évoluer dynamiquement et même de façon antagoniste au fil du temps, mais le changement total est toujours limité. Pour traiter ce cadre, nous proposons un algorithme en ligne efficace basé sur l'itération de valeur des moindres carrés pondérés. Il utilise des poids exponentiels pour oublier doucement les données qui sont loin dans le passé, ce qui pousse l'agent à continuer à explorer pour découvrir les changements. Enfin, au-delà du cadre classique du RL, nous considérons un agent qui interagit avec son environnement sans signal de récompense. Nous proposons d'apprendre une paire de représentations qui mettent en correspondance les paires état-action avec un certain espace latent. Pendant la phase non supervisée, ces représentations sont entraînées en utilisant des interactions sans récompense pour encoder les relations à longue portée entre les états et les actions, via une carte d'occupation prédictive. Au moment du test, lorsqu'une fonction de récompense est révélée, nous montrons que la politique optimale pour cette récompense est directement obtenue à partir de ces représentations, sans aucune planification. Il s'agit d'une étape vers la construction d'agents entièrement contrôlables. Un thème commun de la thèse est la conception d'algorithmes RL prouvables et généralisables. Dans la première et la deuxième partie, nous traitons de la généralisation dans les grands espaces d'états, soit par approximation de fonctions linéaires, soit par agrégation d'états. Dans la dernière partie, nous nous concentrons sur la généralisation sur les fonctions de récompense et nous proposons un cadre d'apprentissage non-supervisé de représentation qui est capable d'optimiser toutes les fonctions de récompense. / Reinforcement Learning (RL) is an agent-oriented learning paradigm concerned with learning by interacting with an uncertain environment. Combined with deep neural networks as function approximators, deep reinforcement learning (Deep RL) allowed recently to tackle highly complex tasks and enable artificial agents to master classic games like Go, play video games from pixels, and solve robotic control tasks. However, a closer look at these remarkable empirical successes reveals some fundamental limitations. First, it has been challenging to combine desirable features of RL algorithms, such as off-policy and multi-step learning with function approximation in a way that leads to both stable and efficient algorithms in large state spaces. Moreover, Deep RL algorithms tend to be very sample inefficient due to the rudimentary exploration-exploitation strategies these approaches employ. Finally, they require an enormous amount of supervised data and end up producing a narrow agent able to solve only the task that it was trained on. In this thesis, we propose novel solutions to the problems of off-policy learning and exploration-exploitation dilemma in large state spaces, as well as self-supervision in RL. On the topic of off-policy learning, we provide two contributions. First, for the problem of policy evaluation, we show that combining popular off-policy and multi-step learning methods with linear value function parameterization could lead to undesirable instability, and we derive a provably convergent variant of these methods. Second, for policy optimization, we propose to stabilize the policy improvement step through an off-policy divergence regularization that constrains the discounted state-action visitation induced by consecutive policies to be close to one another. Next, we study online learning in large state spaces and we focus on two structural assumptions to make the problem tractable: smooth and linear environments. For smooth environments, we propose an efficient online algorithm that actively learns an adaptive partitioning of the joint space by zooming in on more promising and frequently visited regions. For linear environments, we study a more realistic setting, where the environment is now allowed to evolve dynamically and even adversarially over time, but the total change is still bounded. To address this setting, we propose an efficient online algorithm based on weighted least squares value iteration. It uses exponential weights to smoothly forget data that are far in the past, which drives the agent to keep exploring to discover changes. Finally, beyond the classical RL setting, we consider an agent interacting with its environments without a reward signal. We propose to learn a pair of representations that map state-action pairs to some latent space. During the unsupervised phase, these representations are trained using reward-free interactions to encode long-range relationships between states and actions, via a predictive occupancy map. At test time, once a reward function is revealed, we show that the optimal policy for that reward is directly obtained from these representations, with no planning. This is a step towards building fully controllable agents. A common theme in the thesis is the design of provable RL algorithms that generalize. In the first and the second part, we deal with generalization in large state spaces either by linear function approximation or state aggregation. In the last part, we focus on generalization over reward functions and we propose a task-agnostic representation learning framework that is provably able to solve all reward functions.

Cette thèse étudie les communications directes entre les mobiles, appelées communications D2D, en tant que technique prometteuse pour améliorer les futurs réseaux cellulaires. Cette technologie permet une communication directe entre deux terminaux mobiles sans passer par la station de base. La modélisation, l'évaluation et l'optimisation des différents aspects des communications D2D constituent les objectifs fondamentaux de cette thèse et sont réalisés principalement à l'aide des outils mathématiques suivants: la théorie des files d'attente, l'optimisation de Lyapunov et les processus de décision markovien partiellement observable POMDP. Les résultats de cette étude sont présentés en trois parties. Dans la première partie, nous étudions un schéma de sélection entre mode cellulaire et mode D2D. Nous dérivons les régions de stabilité des scénarios suivants: réseaux cellulaires purs et réseaux cellulaires où les communications D2D sont activées. Une comparaison entre ces deux scénarios conduit à l'élaboration d'un algorithme de sélection entre le mode cellulaire et le mode D2D qui permet d'améliorer la capacité du réseau. Dans la deuxième partie, nous développons un algorithme d'allocation de ressources des communications D2D. Les utilisateurs D2D sont en mesure d'estimer leur propre qualité de canal, cependant la station de base a besoin de recevoir des messages de signalisation pour acquérir cette information. Sur la base de cette connaissance disponibles au niveau des utilisateurs D2D, une approche d'allocation des ressources est proposée afin d'améliorer l'efficacité énergétique des communications D2D. La version distribuée de cet algorithme s'avère plus performante que celle centralisée. Dans le schéma distribué des collisions peuvent se produire durant la transmission de l'état des canaux D2D ; ainsi un algorithme de réduction des collisions est élaboré. En outre, la mise en œuvre des algorithmes centralisé et distribué dans un réseau cellulaire, type LTE, est décrite en détails. Dans la troisième partie, nous étudions une politique de sélection des relais D2D mobiles. La mobilité des relais représente un des principaux défis que rencontre toute stratégie de sélection de relais. Le problème est modélisé par un processus contraint de décision markovien partiellement observable qui prend en compte le dynamisme des relais et vise à trouver la politique de sélection de relais qui optimise la performance du réseau cellulaire sous des contraintes de coût. / This thesis considers Device-to-Device (D2D) communications as a promising technique for enhancing future cellular networks. Modeling, evaluating and optimizing D2D features are the fundamental goals of this thesis and are mainly achieved using the following mathematical tools: queuing theory, Lyapunov optimization and Partially Observed Markov Decision Process (POMDP). The findings of this study are presented in three parts. In the first part, we investigate a D2D mode selection scheme. We derive the queuing stability regions of both scenarios: pure cellular networks and D2D-enabled cellular networks. Comparing both scenarios leads us to elaborate a D2D vs cellular mode selection design that improves the capacity of the network. In the second part, we develop a D2D resource allocation algorithm. We observe that D2D users are able to estimate their local Channel State Information (CSI), however the base station needs some signaling exchange to acquire this information. Based on the D2D users' knowledge of their local CSI, we provide an energy efficient resource allocation framework that shows how distributed scheduling outperforms centralized one. In the distributed approach, collisions may occur between the different CSI reporting; thus, we propose a collision reduction algorithm. Moreover, we give a detailed description on how both centralized and distributed algorithms can be implemented in practice. In the third part, we propose a mobile relay selection policy in a D2D relay-aided network. Relays' mobility appears as a crucial challenge for defining the strategy of selecting the optimal D2D relays. The problem is formulated as a constrained POMDP which captures the dynamism of the relays and aims to find the optimal relay selection policy that maximizes the performance of the network under cost constraints.

Cette thèse se penche sur les modèles de choix, des distributions sur des ensembles d'alternatives. Les modèles de choix sur les processus décisionnels de Markov (MDP) peuvent décomposer de très grands espaces alternatifs en procédures étape par étape conçues pour non seulement combattre la malédiction de la dimensionnalité mais aussi pour mieux refléter la dynamique sous-jacente. La première partie est consacrée à l'estimation du temps de trajet dans le cadre de la modélisation du choix de chemin. Les modèles de choix de chemin sont des modèles de choix sur l'ensemble des chemins utilisés pour modéliser le flux de circulation. Intuitivement, le temps de trajet est l'une des caractéristiques les plus importantes lors du choix des chemins, mais les temps de trajet ne sont pas toujours connus. En revanche, le cadre classique suppose que ces deux étapes sont séquentielles, car les temps de trajet des arcs font partie de l'entrée du processus d'estimation du choix de chemin. Pourtant, les interdépendances complexes signifient que ce modèle de choix de chemin peut complémenter toute observation lors de l'estimation des temps de trajet. Nous construisons un modèle statistique pour l'estimation du temps de trajet et proposons de marginaliser les caractéristiques non observées. En utilisant ces idées, nous montrons que nous sommes capables d'apprendre des modèles de choix de chemin sans observer de chemins réels et à différentes granularités. La deuxième partie se concentre sur les échecs des MDP régularisés et comment la régularisation peut avoir des effets secondaires inattendus, tels que la divergence dans les chemins stochastiques les plus courts ou des fonctions de valeur déraisonnablement grandes. Les MDP régularisés ne sont rien d'autre qu'une application des modèles de choix aux MDP. Ils sont utilisés dans l'apprentissage par renforcement (RL) pour obtenir, entre autres choses, un modèle de choix sur les trajectoires possibles pour l'apprentissage par renforcement inverse, transférer des connaissances préalables au modèle, ou obtenir des politiques qui exploitent tous les objectifs dans l'environnement. Ces effets secondaires sont exacerbés dans les espaces d'action dépendants de l'état. Comme mesure d'atténuation, nous introduisons deux transformations potentielles, et nous évaluons leur performance sur un problème de conception de médicaments. / This thesis delves on choice models, distributions on sets of alternatives. Choice models on Markov decision processes (MDPs) can break down very large alternative spaces into step-by-step procedures designed to not only tackle the curse of dimensionality but also to reflect the underlying dynamics better. The first part is devoted to travel time estimation as part of path choice modeling. Path choice models are choice models on the set of paths used to model traffic flow. Intuitively, travel time is one of the more important features when choosing paths, yet travel times are not always known. In contrast, the classical setting assumes that these two steps are sequential, as arc travel times are part of the input of the path choice estimation process. Yet the intricate interdependences mean that that path choice model can complement any observation when estimating travel times. We build a statistical model for travel time estimation and propose marginalizing the unobserved features. Using these ideas, we show that we are able to learn path choice models without observing actual paths and at different granularity. The second part focuses on the failings of regularized MDPs and how regularization may have unexpected side effects, such as divergence in stochastic shortest paths or unreasonably large value functions. Regularized MDPs are nothing but an application of choice models to MDPs. They are used in reinforcement learning (RL) to get, among other things, a choice model on possible trajectories for inverse reinforcement learning, transfer prior knowledge to the model, or to get policies that exploit all goals in the environment. These side effects are exacerbated in state-dependent action spaces. As a mitigation, we introduce two potential transformations, and we benchmark their performance on a drug design problem.

The convergence of various technological and operational advancements has reinstated the interest in On-Demand Air Service (ODAS) as a viable mode of transportation. ODAS enables an end-user to be transported in an aircraft between their desired origin and destination at their preferred time without advance notice. Industry, academia, and the government organizations are collaborating to create technology solutions suited for large-scale implementation of this mode of transportation. Market studies suggest reducing vehicle operating cost per passenger as one of the biggest enablers of this market. To enable ODAS, an ODAS operator controls a fleet of aircraft that are deployed across a set of nodes (e.g., airports, vertiports) to satisfy end-user transportation requests. There is a gap in the literature for a tractable and online methodology that can enable ride-sharing in the on-demand operations while maintaining a publicly acceptable level of service (such as with low waiting time). The need for an approach that not only supports a dynamic-stochastic formulation but can also handle uncertainty with unknowable properties, drives me towards the field of Reinforcement Learning (RL). In this work, a novel two-layer hierarchical RL framework is proposed that can distribute a fleet of aircraft across a nodal network as well as perform real-time scheduling for an ODAS operator. The top layer of the framework - the Fleet Distributor - is modeled as a Partially Observable Markov Decision Process whereas the lower layer - the Trip Request Manager - is modeled as a Semi-Markov Decision Process. This framework is successfully demonstrated and assessed through various studies for a hypothetical ODAS operator in the Chicago region. This approach provides a new way of solving fleet distribution and scheduling problems in aviation. It also bridges the gap between the state-of-the-art RL advancements and node-based transportation network problems. Moreover, this work provides a non-proprietary approach to reasonably model ODAS operations that can be leveraged by researchers and policy makers.