Pour rester performants et compétitifs, les systèmes de production doivent être capables de s’adapter pour faire face aux changements tels que l’évolution de la demande des clients. Il leur est essentiel de pouvoir déterminer quand et comment s’adapter (capacités, etc.). Malheureusement, de tels problèmes sont connus pour être difficiles. Les systèmes de production étant complexes, dynamiques et spécifiques, leurs gestionnaires n’ont pas toujours l’expertise nécessaire ni les prévisions suffisantes concernant l’évolution de leur système. Cette thèse vise à étudier la contribution que peut apporter l’apprentissage automatique à l’auto-adaptation des systèmes de production. Dans un premier temps, nous étudions la façon dont la littérature aborde ce domaine et en proposons un cadre conceptuel dans le but de faciliter l’analyse et la formalisation des problèmes associés. Ensuite, nous étudions une stratégie d’apprentissage à partir de modèles qui ne nécessite pas d’ensemble d’apprentissage. Nous nous intéressons plus précisément à une nouvelle approche basée sur la programmation génétique linéaire visant à extraire des connaissances itérativement à partir d’un modèle de simulation pour déterminer quand et quoi faire évoluer. Notre approche est implémentée à l’aide d’Arena et μGP. Nous l’appliquons à différents exemples qui concernent l’ajout/retrait de cartes dans un système à flux tiré, le déménagement de machines ou encore le changement de politique de réapprovisionnement. Les connaissances qui en sont extraites s’avèrent pertinentes et permettent de déterminer en continu comment chaque système peut s’adapter à des évolutions. De ce fait, elles peuvent contribuer à doter un système d’une forme d’intelligence. Exprimées sous forme d’un arbre de décision, elles sont par ailleurs facilement communicables à un gestionnaire de production. Les résultats obtenus montrent ainsi l’intérêt de notre approche tout en ouvrant de nombreuses voies de recherche. / Manufacturing systems must be able to continuously adapt their characteristics to cope with the different changes that occur along their life, in order to remain efficient and competitive. These changes can take the form of the evolution of customers demand for instance. It is essential for these systems to determine when and how to adapt (e.g., through changes in capacities). Unfortunately, such issues are known to be difficult. As manufacturing systems are complex, dynamic and specific in nature, their managers do not always have all the necessary expertise nor accurate enough forecasts on the evolution of their system. This thesis aims at studying the possible contributions of machine learning to the self-adaptation of manufacturing systems. We first study how the literature deals with self-adaptation and we propose a conceptual framework to facilitate the analysis and the formalization of the associated problems. Then, we study a learning strategy relying on models, which presents the advantage of not requiring any training set. We focus more precisely on a new approach based on linear genetic programming that iteratively extracts knowledge from a simulation model. Our approach is implemented using Arena and μGP. We show its benefits by applying it to increase/decrease the number of cards in a pull control system, to move machines or to change the inventory replenishment policy. The extracted knowledge is found to be relevant for continuously determining how each system can adapt to evolutions. It can therefore contribute to provide these systems with some intelligent capabilities. Moreover, this knowledge is expressed in the simple and understandable form of a decision tree, so that it can also be easily communicated to production managers in view of their everyday use. Our results thus show the interest of our approach while opening many research directions.

This thesis is concerned with the utilization of formal verification techniques in the design of the functional approximations of combinational circuits. We thoroughly study the existing formal approaches for the approximate equivalence checking and their utilization in the approximate circuit development. We present a new method that integrates the formal techniques into the Cartesian Genetic Programming. The key idea of our approach is to employ a new search strategy that drives the evolution towards promptly verifiable candidate solutions. The proposed method was implemented within ABC synthesis tool. Various parameters of the search strategy were examined and the algorithm's performance was evaluated on the functional approximations of multipliers and adders with operand widths up to 32 and 128 bits respectively. Achieved results show an unprecedented scalability of our approach.

This Master's Thesis is focused on the principles of neural networks, primarily convolutional neural networks (CNN). It introduces the evolutionary optimization in the context of neural networks. One of existing libraries devoted to the CNN design was chosen (Keras), analysed and used in image classification tasks. An optimization technique based on cartesian genetic programming that should reduce the complexity of CNN's computation was proposed and implemented. The impact of the proposed technique on CNN behaviour was evaluated in a case study.

This thesis is focused on finding procedures that would accelerate symbolic regressions in Cartesian Genetic Programming. It describes Cartesian Genetic Programming and its use in the task of symbolic regression. It deals with the SIMD architecture and the SSE and AVX instruction set. Several optimizations that lead to a significant acceleration of evolution in Cartesian Genetic Programming are presented. A method of a bit-level parallel simulation that uses AVX2 vectors allows to process 256 input combinations of a logic circuit in paralell. Similarly it is possible to use a byte-level parallel simulation and work with 32 bytes when evolving an image filter. A new method of batch mutation can accelerate the evolution of combinational logic circuits thousand times depending on the problem size. For example, using a combination of these and other methods the evolution of 5 x 5b multipliers took 5.8 seconds on average on an Intel Core i5-4590 processor.

This master thesis focuses on approximate computing applied to image processing. Specifically, the approximation is applied to adaptive thresholding. Two approaches were used, the design of a new system using approximated components and the approximation of an existing algorithm. The resulting effect on thresholding quality was investigated. Experimental evaluation of the first approach shows quality improvements of thresholding with usage of aproximated components. Also, area of found aproximated solutions is smaller. Evaluation of the second approach shows worse quality of thresholding with usage of aproximated components. The second approach is then declared inappropriate.

This thesis focuses on the automation of scenarios creation for Portable Stimulus standard. The main goal of the work is an automatic generation of tests, which are defined as graphs for the Questa inFact tool from the Mentor company. For the automation I used an evolutionary algorithm with using a grammatical evolution.  The generated scenarios are connected to the existing verification environment based on UVM methodology, then the verification of the connected component is started. Based on the achieved functional and structural coverage, the individual's fitness value is calculated and propagated into an evolutionary algorithm.  At the end of the work, experiments are performed on the timer component and the contribution of the proposed evolutionary algorithm is evaluated. The proposed evolutionary algorithm is configurable by  grammar and user-defined basic transactions, which allows a wide range of uses. The evolutionary algorithm managed to achieve high functional and structural coverage on the verified timer component.

This Master´s thesis implements computer program in Java, useful for automatic model generating, specially in symbolic regression problem. Thesis includes short description of genetic programming (GP) and own implementation with advanced GP operands (non-destructive operations, elitism, exptression reduction). Mathematical model is generating by symbolic regression, exacly for choosen data set. For functioning check are used test tasks. Optimal settings is found for choosen GP parameters.

Disertační práce je zaměřena na optimalizaci průběhu pracovních operací v logistických skladech a distribučních centrech. Hlavním cílem je optimalizovat procesy plánování, rozvrhování a odbavování. Jelikož jde o problém patřící do třídy složitosti NP-težký, je výpočetně velmi náročné nalézt optimální řešení. Motivací pro řešení této práce je vyplnění pomyslné mezery mezi metodami zkoumanými na vědecké a akademické půdě a metodami používanými v produkčních komerčních prostředích. Jádro optimalizačního algoritmu je založeno na základě genetického programování řízeného bezkontextovou gramatikou. Hlavním přínosem této práce je a) navrhnout nový optimalizační algoritmus, který respektuje následující optimalizační podmínky: celkový čas zpracování, využití zdrojů, a zahlcení skladových uliček, které může nastat během zpracování úkolů, b) analyzovat historická data z provozu skladu a vyvinout sadu testovacích příkladů, které mohou sloužit jako referenční výsledky pro další výzkum, a dále c) pokusit se předčit stanovené referenční výsledky dosažené kvalifikovaným a trénovaným operačním manažerem jednoho z největších skladů ve střední Evropě.

This thesis deals with employing coevolutionary principles to the image filter design. Evolutionary algorithms are very advisable method for image filter design. Using coevolution, we can add the processes, which can accelerate the convergence by interactions of candidate filters population with population of fitness predictors. Fitness predictor is a small subset of the training set and it is used to approximate the fitness of the candidate solutions. In this thesis, indirect encoding is used for predictors evolution. This encoding represents a mathematical expression, which selects training vectors for candidate filters fitness prediction. This approach was experimentally evaluated in the task of image filters for various intensity of random impulse and salt and pepper noise design and the design of the edge detectors. It was shown, that this approach leads to adapting the number of target objective vectors for a particular task, which leads to computational complexity reduction.

In recent years, there has been a strong need for the design of integrated  circuits showing low power consumption. It is possible to create intentionally approximate circuits which don't fully implement the specified logic behaviour, but exhibit improvements in term of area, delay and power consumption. These circuits can be used in many error resilient applications, especially in signal and image processing, computer graphics, computer vision and machine learning. This work describes an evolutionary approach to approximate design of arithmetic circuits and other more complex systems. This text presents a parallel calculation of a fitness function. The proposed method accelerated evaluation of 8-bit approximate multiplier 170 times in comparison with the common version. Evolved approximate circuits were used in different types of edge detectors.