A utilização de sistemas de controle para estabilidade de conjuntos de barras para pulverizadores agrícolas é uma tendência devida principalmente aos problemas ambientais e de custo. Neste trabalho, o comportamento dinâmico de um pulverizador de arrasto é analisado através de um modelo não linear, obtido através de técnicas de modelagem de sistemas multicorpos utilizando-se o programa ADAMS. Foram utilizadas duas estratégias de controle PID e \"fuzzy\" a partir de medidas obtidas com fusão de sensores. A estratégia de controle clássica PID foi desenvolvida e implementada no modelo não linear no ADAMS através de ferramentas internas existentes no programa. Já a estratégia \"fuzzy\" foi desenvolvida e implementada no modelo não linear no ADAMS através da técnica de co-simulação ADAMS/Matlab. O comportamento dos sistemas de controle foi investigado através de simulação computacional. Foram testados alguns tipos de entradas (entrada degrau, entrada harmônica, entrada randômica e entrada randômica com descontinuidades bruscas). Em todas as simulações os resultados obtidos com os sistemas de controles ativos mostraram melhor estabilidade do conjunto de barras. Entre as leis de controle implementadas (PID e \"fuzzy\") não houve grandes diferenças entre as oscilações da barra exceto na entrada do tipo randômica com descontinuidades bruscas. Neste caso a lei de controle \"fuzzy\" apresentou uma grande melhoria com boa atenuação das oscilações do conjunto de barras quando comparadas com a aplicação do sistema de controle PID. / The use of active control systems for stability of booms in agricultural sprayers trend is mainly due to the environmental and costs question. In this work, the dynamic behavior of a trailed sprayer is analyzed using a nonlinear model, obtained through techniques of modeling multibody systems using the ADAMS. It is used two active control strategies, PID classical control and fuzzy, with measured data from sensor fusion. The classical PID control strategy was developed and implemented in a nonlinear model on ADAMS software using existing tools built into the program. Fuzzy was another strategy developed and implemented in the nonlinear model on ADAMS software using a technique of co-simulation ADAMS/Matlab. The behavior of control systems was investigated through computer simulation. It was tested some types of inputs (step input, harmonic input, random input and random input with abrupt discontinuities). All simulations data obtained from the applications of active systems showed better stability for boom assembly. Among the implemented two active control laws (PID and \"fuzzy\") there were no significant differences between the oscillations attenuation of the boom, except with the random input with abrupt discontinuities. wherein this case the application of the active control \"fuzzy\" strategy developed better stability on boom than the application of PID control.

Energieeffizienz hat in Forschung und Alltag eine zentrale Bedeutung. Arbeiten verschiedene elektronische Geräte zusammen, um gemeinsam eine Regelungsaufgabe zu lösen, müssen sie miteinander kommunizieren. Ein Beispiel aus dem Alltag sind Funk-Raumtemperaturregler, bei denen ein batteriebetriebener Temperatursensor und ein Heizungsaktor (Stellantrieb am Heizungsventil) über drahtlose Kommunikation zusammenarbeiten. Diese Kommunikation benötigt oft mehr Energie als der Betrieb der eigentlichen (elektronischen) Funktionalität der Teilsysteme. Energieeffizienter als die in Regelkreisen übliche periodische (äquidistante) Abtastung ist – durch eine Verringerung der Nachrichtenrate – eine ereignisbasierte Abtastung. Send-on-Delta-Abtastung ist die am weitesten verbreitete Art der ereignisbasierten Abtastung. Dabei wird der Wert der Regelgröße (im Beispiel die Raumtemperatur) nicht in konstanten Zeitintervallen übertragen, sondern nur dann, wenn er sich um einen bestimmten Betrag geändert hat. Der mit einem Anteil von über 90 % im Praxiseinsatz am weitesten verbreitete Reglertyp ist der PID-Regler, wobei die meisten als „PID-Regler“ bezeichneten Regler aus verschiedenen Gründen keinen D-Anteil (Differential-Anteil) verwenden und daher als „PI-Regler“ bezeichnet werden können. Die vorliegende Arbeit verfolgt das Ziel, systematisch zu untersuchen, wie man PI-Regler einstellen sollte, um neben dem Erreichen einer hohen Regelgüte auch die Vorteile der Send-on-Delta-Abtastung bezüglich der Netzlastreduktion und Energieeffizienz bestmöglich auszunutzen. Die „Gewichtung“ dieser sich teilweise widersprechenden Kriterien ist anwendungsspezifisch einstellbar. / Energy efficiency is very important both in science and everyday life. If different electronic devices work together, for example for solving a control task together, they have to communicate with each other. An everyday life example are room temperature controllers using radio communication between a battery-powered temperature sensor and a heating actuator. This communication often needs more energy than the operation of the actual (electronic) functionality of the components. More energy-efficient than the commonly used periodic sampling is event-based sampling, due to the reduction of the message rate. Send-on-delta sampling is the most widely-known kind of event-based sampling. In that case, the value of the controlled variable (e.g. the room temperature) is not transmitted equidistantly but only when it has changed by a specific amount. The most successful controller in practice is the PID controller. The most so-called “PID controllers” do not use the D part (differential action) for several reasons and can therefore be called “PI controllers”. This work analyses systematically how the parameters of a PI controller should be tuned to reach besides a high control quality also a good exploitation of the advantages of send-on-delta sampling regarding network load reduction and energy efficiency. The “weighting” of these partially contradicting criteria is application specifically adjustable.

L’équipe Physique des Galaxies du Laboratoire d’Astrophysique de Marseille a développé un nouveau type d'interféromètre de Fabry-Perot, équipé de trois actionneurs piézoélectriques amplifiés et de trois capteurs capacitifs permettant le contrôle de l'espacement et du parallélisme des lames de verres de l'ordre de 200 µm avec une précision de positionnement du nm.L'objectif visé de ce manuscrit, composé de 3 parties, est le pilotage de cet interféromètre. La première partie, composée des chapitres 1 et 2, présente les généralités sur l'interférométrie de Fabry-Perot puis décrit les instruments 3DNTT et BTFI où seront installés l'interféromètre de nouvelle génération et son contrôleur associé. La conception et la réalisation d'un amplificateur hybride en vue de la réduction des non-linéarités des actionneurs piézoélectriques mettent fin à cette partie.La deuxième partie, chapitres 3 et 4, décrit le développement et la réalisation du contrôleur. Pour cela, une démarche basée sur le concept du Co-design a été adoptée.Le contrôleur ainsi réalisé est composé d'une carte de développement Microzed dont le cœur est un système sur puce de la série Zynq 7000 EPP et d'une carte d'interfaçage comportant des convertisseurs 3 ADC et 3 DAC et des circuits d'alimentation. La troisième et dernière partie, chapitres 5 et 6, traite de la modélisation d'état de l’interféromètre de Fabry-Perot et de son contrôle : un contrôle classique basé sur la régulation PID et un contrôle robuste et optimal basé sur le filtrage de KALMAN. Cette dernière partie conclut sur les perspectives pouvant découler des contributions de ce travail sur le contrôle et la commande Fabry-Perot. / The Physics of Galaxies Team of Laboratoire d'Astrophysique de Marseille (LAM) has developed a new type of Fabry-Perot, with three amplified piezoelectric actuators and three capacitive sensors to control the spacing and parallelism of mirror plates of approximately 200 µm with a positioning accuracy of 3 nm.The purpose of this manuscript, consisting of 3 parts is the control of this interferometer.The first part, consisting of Chapters 1 and 2 presents the general interferometry Fabry-Perot and then describes 3DNTT and BTFI instruments which will be installed the next generation interferometer and its associated controller. The design and implementation of a hybrid amplifier to reduce non-linearities of the piezoelectric actuators (hysteresis and creep) end this first part.The second part, Chapters 3 and 4, describes the development and implementation of the controller.For this, an approach based on codesign concept was adopted. The thus achieved controller consists of a Microzed development board whose heart is a system on chip of the 7000 series Zynq EPP (FPGA + Dual-Core ARM Cortex A9) and an interface card with converters (3 ADC and 3 DAC) and power supply circuits.For the finalization of the controller, two steps are necessary: hardware design in Xilinx Vivado and software design in Xilinx SDK.The third and final section, chapters 5 and 6 deals with the Fabry-Perot space-state modeling and its control: a classic control based on PID control and a robust and optimal control based on KALMAN filtering. This last part concludes the outlook may result from contributions of this work on the monitoring and control of the Fabry-Perot.

The aim of the thesis was to implement and compare control methods in a model of a main battle tank. Three controllers were implemented in a two axis gimbal model and their performances were compared. The comparisons were performed using step changes in the reference signal, frequency analysis of an oscillating reference signal and disturbances, and turret mass uncertainties. The results showed that the sliding mode controller had the best performance for both reference changes and disturbance attenuation. The PID controller had a better performance for the change in reference, compared to the model predictive controller, but a significantly worse disturbance attenuation. Due to model approximations, such as assuming ideal engines and noise reduction, the results likely show a better performance than what can be expected if applied on a real main battle tank. Therefore, the results show an upper limit of the stabilization performance of turret and barrel control and should only be used to compare the controllers. / Målet med uppsatsen var att implementera och jämföra reglermetoder för en teoretisk modell av en modern stridsvagn. Tre metoder implementerades i ett tvåaxligt gimbalsystem och deras prestanda utvärderades. Mer specifikt utvärderades regulatorernas respons för en referensändring, en referensstörning och osäkerheter i tornets massa och massfördelning. Utifrån dessa resultat jämfördes sedan reglermetoderna. Resultaten visade att sliding mode regulatorn hade bäst prestanda för både referensändring och referensstörningen när man tittar på frekvensanalys av mät- och processstörningar. PID-regulatorn hade en bättre prestanda än MPC-regulatorn för en referensändring men en sämre respons för en referensstö-rning. På grund av modellförenklingar som till exempel antaganden om ideala motorer och brusreduktion visar resultaten troligen en något bättre prestanda än vad som kan uppnås på en riktigt stridsvagn. Det medför att resultaten visar en övre gräns för vad styrsystemet för en stridsvagns torn- och eldrörsrotation kan uppnå.

This thesis project has been conducted during a five-month research exchange visit to the Space Structure Dynamics and Control research group at University College Dublin. This report presents the design, development, and validation of a nanosatellite attitude control testbed. The testbed was designed to replicate the microgravity conditions of space by utilising an air bearing, enabling single-axis rotational motion for a 1U CubeSat-sized nanosatellite. The novel aspect of this research is the inclusion of two-degree-of-freedom, lumped-mass flexible appendages on either side of the nanosatellite, emulating a lightweight, flexible space structure. These flexible appendages were designed based on the stiffness characteristics of a deployable CubeSat solar array system found in existing literature, with exaggerated motion to amplify the measurable effects of various control approaches. The central focus of this project was the development of an avionics stack closely resembling CubeSat attitude control boards. The stack uses an STM32 microcontroller as the primary attitude control computer, and a suite of off the shelf breakout boards for sensors and wireless telemetry systems. Power, serial and I2C buses connect the attitude control board and the onboard computer board. A reaction wheel actuator controls the Euler heading attitude. The testbed was designed as an experimental platform for validating control algorithms developed through a model-based approach. Integration with the Simulink Embedded Coder toolbox allows for the compilation of Simulink models into C code, facilitating direct execution on the testbed. The testbed’s physical construction involves 3D printed ABS components, with the inclusion of load cells to measure disturbance torques from the excited flexible appendages. Results from validation experiments show that a simple PID controller causes significant excitation in the flexible appendages during a slew manoeuvre. However, the introduction of an input shaped attitude profile tailored to the natural frequency of the appendages successfully reduced the measured appendage excitation by 50%. Conversely, the force impedance wave based control approach did not show a reduction in appendage excitation, but shows promise for further developments in future work. In conclusion, the testbed has successfully achieved its predefined project objectives, albeit requiring further refinement, particularly in the telemetry down-link system. It is recommended that future work focuses on enhancement of the telemetry system, and validation of a model based approach to controller design. / Detta examensarbete har utförts under ett fem månaders forskningsutbyte vid forskningsgruppen Space Structure Dynamics and Control vid University College Dublin. Denna rapport presenterar design, utveckling och validering av en testbädd för attitydstyrning av en nanosatellit. Testbädden utformades för att efterlikna mikrogravitationsförhållandena i rymden genom att använda ett luftbärande lager, vilket möjliggör rotationsrörelse kring en axel för en nanosatellit av storleken 1U CubeSat. En unik aspekt av detta arbete är inkluderingen av två-frihet-graders, sammansatta flexibla påhängen på vardera sidan av nanosatelliten, vilket efterliknar en lätt, flexibel rymdstruktur. Dessa flexibla påhängen utformades med utgångspunkt från styvhetsegenskaperna hos ett utskjutbart CubeSat-solcellssystem som finns i befintlig litteratur, med överdriven rörelse för att förstärka de mätbara effekterna av olika styrmetoder. Det centrala fokuset för detta projekt var utvecklingen av en avionikstack som nära liknar CubeSat-attitydregleringkort. Stacken använder en STM32-mikrokontroller som primär dator för attitydkontroll och ett urval färdiga breakout-kort för sensorer och trådlös telemetri. Strömförsörjning, serie- och I2C-bussar ansluter attitydregleringkortet och omborddatorn med en reaktionshjulsaktuator som styr Euler-attityd. Testbädden utformades som en experimentell plattform för att validera styralgoritmer som utvecklats med hjälp av en modellbaserad metod. Integration med Simulink Embedded Coder möjliggör kompilering av Simulink-modeller till C-kod, vilket underlättar direkt exekvering på testbädden. Testbäddens fysiska konstruktion innefattar 3D-utskrivna ABS-komponenter med inkludering av lastceller för att mäta störningsmoment från de exciterade flexibla påhängen. Resultaten från valideringsexperiment visar att en enkel PID-regulator orsakar betydande excitation i de flexibla påhängena under en vridningsmanöver. Dock lyckades en input-formad attitydprofil som anpassats till de flexibla påhängenas naturliga frekvens framgångsrikt minska den uppmätta excitationen med 50%. Omvänt visade den kraftimpedansvågbaserade styrmetoden inte någon minskning i excitation, men visar potential för vidare utvecklingar i framtida arbete. Sammanfattningsvis har testbädden framgångsrikt uppnått sina fördefinierade projektmål, om än med behov av ytterligare förfining, särskilt i telemetrisystemet. Det rekommenderas att framtida arbete fokuserar på förbättring av telemetrisystemet och validering av en modellbaserad ansats till styrdesign.

Tato diplomová práce je zaměřena na vývoj simulačního prostředí v Matlab/Simulink zvoleného letadla ve známém letovém režimu. Pozice a orientace letadla pohybujícího se ve vzduchu je popsána pohybovými rovnicemi se šesti stup\v{n}i volnosti. Soustava translačních, rotačních a kinematických rovnic tvoří soustavu devíti nelineárních diferenciálních rovnic prvního řádu. Tyto rovnice lze linearizovat okolo nějakého rovnovážného stavu, který budeme nazývat letovým režimem. Součástí simulačního prostředí je řídící systém letadla založený na PID regulaci. Základem je návrh autopilota, který řídí úhel podélného sklonu a úhel příčného náklonu. Součástí návrhu jsou takzvané „flight director\textquotedblright \phantom{s}m\'dy jako udržení výšky, volba kursu, regulace vertikální rychlosti, změna výšky, zachycení požadované výšky a navigační m\'{o}d založený na nelineárním navigačním zákonu. Optimalizace regulátorů za použití PSO algoritmu a Pareto optimalitě je využita pro nastavení parametrů PID regulátoru. Simulační prostředí je vizualizováno v softwaru FlightGear.

Energieeffizienz hat in Forschung und Alltag eine zentrale Bedeutung. Arbeiten verschiedene elektronische Geräte zusammen, um gemeinsam eine Regelungsaufgabe zu lösen, müssen sie miteinander kommunizieren. Ein Beispiel aus dem Alltag sind Funk-Raumtemperaturregler, bei denen ein batteriebetriebener Temperatursensor und ein Heizungsaktor (Stellantrieb am Heizungsventil) über drahtlose Kommunikation zusammenarbeiten. Diese Kommunikation benötigt oft mehr Energie als der Betrieb der eigentlichen (elektronischen) Funktionalität der Teilsysteme. Energieeffizienter als die in Regelkreisen übliche periodische (äquidistante) Abtastung ist – durch eine Verringerung der Nachrichtenrate – eine ereignisbasierte Abtastung. Send-on-Delta-Abtastung ist die am weitesten verbreitete Art der ereignisbasierten Abtastung. Dabei wird der Wert der Regelgröße (im Beispiel die Raumtemperatur) nicht in konstanten Zeitintervallen übertragen, sondern nur dann, wenn er sich um einen bestimmten Betrag geändert hat. Der mit einem Anteil von über 90 % im Praxiseinsatz am weitesten verbreitete Reglertyp ist der PID-Regler, wobei die meisten als „PID-Regler“ bezeichneten Regler aus verschiedenen Gründen keinen D-Anteil (Differential-Anteil) verwenden und daher als „PI-Regler“ bezeichnet werden können. Die vorliegende Arbeit verfolgt das Ziel, systematisch zu untersuchen, wie man PI-Regler einstellen sollte, um neben dem Erreichen einer hohen Regelgüte auch die Vorteile der Send-on-Delta-Abtastung bezüglich der Netzlastreduktion und Energieeffizienz bestmöglich auszunutzen. Die „Gewichtung“ dieser sich teilweise widersprechenden Kriterien ist anwendungsspezifisch einstellbar. / Energy efficiency is very important both in science and everyday life. If different electronic devices work together, for example for solving a control task together, they have to communicate with each other. An everyday life example are room temperature controllers using radio communication between a battery-powered temperature sensor and a heating actuator. This communication often needs more energy than the operation of the actual (electronic) functionality of the components. More energy-efficient than the commonly used periodic sampling is event-based sampling, due to the reduction of the message rate. Send-on-delta sampling is the most widely-known kind of event-based sampling. In that case, the value of the controlled variable (e.g. the room temperature) is not transmitted equidistantly but only when it has changed by a specific amount. The most successful controller in practice is the PID controller. The most so-called “PID controllers” do not use the D part (differential action) for several reasons and can therefore be called “PI controllers”. This work analyses systematically how the parameters of a PI controller should be tuned to reach besides a high control quality also a good exploitation of the advantages of send-on-delta sampling regarding network load reduction and energy efficiency. The “weighting” of these partially contradicting criteria is application specifically adjustable.

This master thesis aims to develop a test methodology for electromechanical linear actuators. A linear actuator acts as a linear motor, converting a power source to linear motion. The electromechanical linear actuator in this project has an electric motor as its power source and uses a rack and pinion system to transfer that power to linear motion.  The test methodology is to impose a force onto the rack of the actuator, to ensure that operation under a load scenario is within specification. To accomplish this, the design of a test rig implementation is analyzed. The test rig consists of the test unit, which is to be tested, the load unit, which is to provide the load force, and a control system for the load unit. The load unit is another linear actuator and is controlled via a load cell. The load cell gives out the load force being applied, and the controller gives out the corresponding appropriate motor command to the load unit to ensure the load force is as desired. This analysis is done through simulation of the setup. Viable options for the setup were first analyzed in order to implement the deemed promising options for a setup into a simulation environment. The simulation environment in this project was Simscape, an extension of MATLAB’s Simulink. In simulation the parameters for the test rig were rigorously analyzed, in order to determine acceptable thresholds. The primary load unit tested was another electromechanical linear actuator from Cascade Drives, the model A-100-8P. Two secondary setups, one using the same model as being tested, and another setup using two of the models being tested. Simulation found that the suggested options’ applied load force have a poor rise time, large overshoot and substantial oscillation errors. The primary source for this was determined to be the latency between load cell input, and motor command output in the controller. The poor metrics from the result could lead to problems when emergency braking, and with a long honing period, which would render most test data unusable.

Le déploiement généralisé de modèles d’apprentissage automatique de plus en plus performants a entraîné des pressions croissantes pour améliorer la robustesse, la sécurité et l’équité de ces modèles—-souvent en raison de considérations réglementaires et éthiques. En outre, la mise en œuvre de solutions d’intelligence artificielle dans des applications réelles est limitée par leur incapacité actuelle à garantir la conformité aux normes industrielles et aux réglementations gouvernementales. Les pipelines standards pour le développement de modèles d’apprentissage automatique adoptent une mentalité de “construire maintenant, réparer plus tard”, intégrant des mesures de sécurité a posteriori. Cette accumulation continue de dette technique entrave le progrès du domaine à long terme. L’optimisation sous contraintes offre un cadre conceptuel accompagné d’outils algorithmiques permettant d’imposer de manière fiable des propriétés complexes sur des modèles d’apprentissage automatique. Cette thèse appelle à un changement de paradigme dans lequel les contraintes constituent une partie intégrante du processus de développement des modèles, visant à produire des modèles d’apprentissage automatique qui sont intrinsèquement sécurisés par conception. Cette thèse offre une perspective holistique sur l’usage de l’optimisation sous contraintes dans les tâches d’apprentissage profond. Nous examinerons i) la nécessité de formulations contraintes, ii) les avantages offerts par le point de vue de l’optimisation sous contraintes, et iii) les défis algorithmiques qui surgissent dans la résolution de ces problèmes. Nous présentons plusieurs études de cas illustrant l’application des techniques d’optimisation sous contraintes à des problèmes courants d’apprentissage automatique. Dans la Contribution I, nous plaidons en faveur de l’utilisation des formulations sous contraintes en apprentissage automatique. Nous soutenons qu’il est préférable de gérer des régularisateurs interprétables via des contraintes explicites plutôt que par des pénalités additives, particulièrement lorsqu’il s’agit de modèles non convexes. Nous considérons l’entraînement de modèles creux avec une régularisation L0 et démontrons que i) il est possible de trouver des solutions réalisables et performantes à des problèmes de grande envergure avec des contraintes non convexes ; et que ii) l’approche contrainte peut éviter les coûteux ajustements par essais et erreurs inhérents aux techniques basées sur les pénalités. La Contribution II approfondit la contribution précédente en imposant des contraintes explicites sur le taux de compression atteint par les Représentations Neuronales Implicites—-une classe de modèles visant à entreposer efficacement des données (telles qu’une image) dans les paramètres d’un réseau neuronal. Dans ce travail, nous nous concentrons sur l’interaction entre la taille du modèle, sa capacité représentationnelle, et le temps d’entraînement requis. Plutôt que de restreindre la taille du modèle à un budget fixe (qui se conforme au taux de compression requis), nous entraînons un modèle surparamétré et creux avec des contraintes de taux de compression. Cela nous permet d’exploiter la puissance de modèles plus grands pour obtenir de meilleures reconstructions, plus rapidement, sans avoir à nous engager à leur taux de compression indésirable. La Contribution III présente les avantages des formulations sous contraintes dans une application réaliste de la parcimonie des modèles avec des contraintes liées à l’équité non différentiables. Les performances des réseaux neuronaux élagués se dégradent de manière inégale entre les sous-groupes de données, nécessitant ainsi l’utilisation de techniques d’atténuation. Nous proposons une formulation qui impose des contraintes sur les changements de précision du modèle dans chaque sous-groupe, contrairement aux travaux antérieurs qui considèrent des contraintes basées sur des métriques de substitution (telles que la perte du sous-groupe). Nous abordons les défis de la non-différentiabilité et de la stochasticité posés par nos contraintes proposées, et démontrons que notre méthode s’adapte de manière fiable aux problèmes d’optimisation impliquant de grands modèles et des centaines de sous-groupes. Dans la Contribution IV, nous nous concentrons sur la dynamique de l’optimisation lagrangienne basée sur le gradient, une technique populaire pour résoudre les problèmes sous contraintes non convexes en apprentissage profond. La nature adversariale du jeu min-max lagrangien le rend sujet à des comportements oscillatoires ou instables. En nous basant sur des idées tirées de la littérature sur les régulateurs PID, nous proposons un algorithme pour modifier les multiplicateurs de Lagrange qui offre une dynamique d’entraînement robuste et stable. Cette contribution met en place les bases pour que les praticiens adoptent et mettent en œuvre des approches sous contraintes avec confiance dans diverses applications réelles. Dans la Contribution V, nous fournissons un aperçu de Cooper : une bibliothèque pour l’optimisation sous contraintes basée sur le lagrangien dans PyTorch. Cette bibliothèque open-source implémente toutes les contributions principales présentées dans les chapitres précédents et s’intègre harmonieusement dans le cadre PyTorch. Nous avons développé Cooper dans le but de rendre les techniques d’optimisation sous contraintes facilement accessibles aux chercheurs et praticiens de l’apprentissage automatique. / The widespread deployment of increasingly capable machine learning models has resulted in mounting pressures to enhance the robustness, safety and fairness of such models--often arising from regulatory and ethical considerations. Further, the implementation of artificial intelligence solutions in real-world applications is limited by their current inability to guarantee compliance with industry standards and governmental regulations. Current standard pipelines for developing machine learning models embrace a “build now, fix later” mentality, retrofitting safety measures as afterthoughts. This continuous incurrence of technical debt hinders the progress of the field in the long-term. Constrained optimization offers a conceptual framework accompanied by algorithmic tools for reliably enforcing complex properties on machine learning models. This thesis calls for a paradigm shift in which constraints constitute an integral part of the model development process, aiming to produce machine learning models that are inherently secure by design. This thesis provides a holistic perspective on the use of constrained optimization in deep learning tasks. We shall explore i) the need for constrained formulations, ii) the advantages afforded by the constrained optimization standpoint and iii) the algorithmic challenges arising in the solution of such problems. We present several case-studies illustrating the application of constrained optimization techniques to popular machine learning problems. In Contribution I, we advocate for the use of constrained formulations in machine learning. We argue that it is preferable to handle interpretable regularizers via explicit constraints, rather than using additive penalties, specially when dealing with non-convex models. We consider the training of sparse models with L0-regularization and demonstrate that i) it is possible to find feasible, well-performing solutions to large-scale problems with non-convex constraints; and that ii) the constrained approach can avoid the costly trial-and-error tuning inherent to penalty-based techniques. Contribution II expands on the previous contribution by imposing explicit constraints on the compression-rate achieved by Implicit Neural Representations—-a class of models that aim to efficiently store data (such as an image) within a neural network’s parameters. In this work we concentrate on the interplay between the model size, its representational capacity and the required training time. Rather than restricting the model size to a fixed budget (that complies with the required compression rate), we train an overparametrized, sparse model with compression-rate constraints. This allows us to exploit the power of larger models to achieve better reconstructions, faster; without having to commit to their undesirable compression rate. Contribution III showcases the advantages of constrained formulations in a realistic model sparsity application with non-differentiable fairness-related constraints. The performance of pruned neural networks degrades unevenly across data sub-groups, thus requiring the use of mitigation techniques. We propose a formulation that imposes constraints on changes in the model accuracy in each sub-group, in contrast to prior work which considers constraints based on surrogate metrics (such as the sub-group loss). We address the non-differentiability and stochasticity challenges posed by our proposed constraints, and demonstrate that our method scales reliably to optimization problems involving large models and hundreds of sub-groups. In Contribution IV, we focus on the dynamics of gradient-based Lagrangian optimization, a popular technique for solving the non-convex constrained problems arising in deep learning. The adversarial nature of the min-max Lagrangian game makes it prone to oscillatory or unstable behaviors. Based on ideas from the PID control literature, we propose an algorithm for updating the Lagrange multipliers which yields robust, stable training dynamics. This contribution lays the groundwork for practitioners to adopt and implement constrained approaches confidently in diverse real-world applications. In Contribution V, we provide an overview of Cooper: a library for Lagrangian-based constrained optimization in PyTorch. This open-source library implements all the core contributions presented in the preceding chapters and integrates seamlessly with the PyTorch framework. We developed Cooper with the goal of making constrained optimization techniques readily available to machine learning researchers and practitioners.