In this thesis direct numerical simulations are used to investigate two phenomenain shear flows: laminar-turbulent transition over a flat plate and periodicvortex shedding induced by a jet in cross flow. The emphasis is on understanding and controlling the flow dynamics using tools from dynamical systems and control theory. In particular, the global behavior of complex flows is describedand low-dimensional models suitable for control design are developed; this isdone by decomposing the flow into global modes determined from spectral analysisof various linear operators associated with the Navier–Stokes equations.Two distinct self-sustained global oscillations, associated with the sheddingof vortices, are identified from direct numerical simulations of the jet incrossflow. The investigation is split into a linear stability analysis of the steadyflow and a nonlinear analysis of the unsteady flow. The eigenmodes of theNavier–Stokes equations, linearized about an unstable steady solution revealthe presence of elliptic, Kelvin-Helmholtz and von K´arm´an type instabilities.The unsteady nonlinear dynamics is decomposed into a sequence of Koopmanmodes, determined from the spectral analysis of the Koopman operator. Thesemodes represent spatial structures with periodic behavior in time. A shearlayermode and a wall mode are identified, corresponding to high-frequency andlow-frequency self-sustained oscillations in the jet in crossflow, respectively.The knowledge of global modes is also useful for transition control, wherethe objective is to reduce the growth of small-amplitude disturbances to delaythe transition to turbulence. Using a particular basis of global modes, knownas balanced modes, low-dimensional models that capture the behavior betweenactuator and sensor signals in a flat-plate boundary layer are constructed andused to design optimal feedback controllers. It is shown that by using controltheory in combination with sensing/actuation in small, localized, regionsnear the rigid wall, the energy of disturbances may be reduced by an order of magnitude.

Concepts in optimal search theory have been used in human-based aerial search since World War II. This thesis addresses the technical and theoretical issues necessary to apply this crucial theory to search path planning for Small Unmanned Aerial Vehicles (SUAVs). A typical search often requires that more than one target be located. Accordingly, a method is presented to locate multiple targets in three dimensions, as well as to differentiate between them. However, significant error can be present when locating targets from an airborne platform, and the idea of target quality is also introduced as a way to describe the reliability of target estimates. Flight test results are presented to validate the target differentiation algorithm. In this test, five out of six targets as close as 6.1 m apart are located and differentiated with less than four meters of error. This flight test also provides color information that is useful in generating artificial target images and understanding the target detection probability. Image skew is then incorporated into the detection probability model, and a function is derived that predicts target detection as a function of distance. In order to measure the effectiveness of search algorithms with this model, the concept of a probability map is introduced. This map can be updated as the search progresses, and is stored on a probability grid containing nodes that keep track of the probable target locations and the probability of detection. Using this tool, a search width is developed for a given airborne agent. The search width is then used to derive optimal search performance based on a given probability map and SUAV. Finally, the concepts of efficiency and completeness are given specific definitions in the context of discrete search. These metrics are used to develop a search plan that focuses on efficiency, and one that focuses on completeness. Example simulations are used to illustrate the conditions under which each plan might be desirable, and a composite search strategy is presented that combines both plans.

The research presented in this thesis was conducted under the joint supervision of Pierre-Luc Bacon, affiliated with Mila - Quebec Artificial Intelligence Institute and Université de Montréal, and Ross Goroshin, affiliated with Google DeepMind. The involvement of both supervisors was integral to the development and completion of this work. / Cette thèse est ancrée dans deux aspirations principales: (i) l'extension des longueurs de séquence pour une fidélité de prédiction supérieure pendant les phases d'entraînement et de test, et (ii) l'amélioration de l'efficacité computationnelle des modèles de séquence. Le défi fondamental de la modélisation de séquences réside dans la prédiction ou la génération précise sur de longs horizons. Les modèles traditionnels, tels que les Réseaux Neuronaux Récurrents (RNN), possèdent des capacités intrinsèques pour la gestion de séquences, mais présentent des lacunes sur de longues séquences. Le premier article, "Correction de Cours des Représentations de Koopman," introduit le Réencodage Périodique pour les Autoencodeurs de Koopman, offrant une solution à la dérive dans les prédictions à long horizon, assurant la stabilité du modèle sur de longues séquences. Les défis subséquents des RNN ont orienté l'attention vers les Transformateurs, avec une longueur de contexte bornée et un temps d'exécution quadratique. Des innovations récentes dans les Modèles d'Espace d'État (SSM) soulignent leur potentiel pour la modélisation de séquences. Notre second article, "Transformateurs d'État-Block," exploite les puissantes capacités de contextualisation des SSM, fusionnant les forces des Transformateurs avec les avantages des SSM. Cette fusion renforce la modélisation linguistique, surtout dans les contextes exigeant une large inference et contexte. En essence, cette thèse se concentre sur l'avancement de l'inférence de séquence à longue portée, chaque article offrant des approches distinctes pour améliorer la portée et la précision de la modélisation prédictive dans les séquences, incarnées par le titre "Au-delà de l'Horizon." / This thesis is anchored in two principal aspirations: (i) the extension of sequence lengths for superior prediction fidelity during both training and test phases, and (ii) the enhancement of computational efficiency in sequence models. The fundamental challenge in sequence modeling lies in accurate prediction or generation across extended horizons. Traditional models, like Recurrent Neural Networks (RNNs), possess inherent capacities for sequence management, but exhibit shortcomings over extended sequences. The first article, "Course Correcting Koopman Representations," introduces Periodic Reencoding for Koopman Autoencoders, offering a solution to the drift in long-horizon predictions, ensuring model stability across lengthy sequences. Subsequent challenges in RNNs have shifted focus to Transformers, with a bounded context length and quadratic runtime. Recent innovations in State-Space Models (SSMs) underscore their potential for sequence modeling. Our second article, "Block-State Transformers," exploits the potent contextualization capabilities of SSMs, melding Transformer strengths with SSM benefits. This fusion augments language modeling, especially in contexts demanding extensive range inference and context. In essence, this thesis revolves around advancing long-range sequence inference, with each article providing distinctive approaches to enhance the reach and accuracy of predictive modeling in sequences, epitomized by the title "Beyond the Horizon."