Ce travail concerne l’étude d’un réacteur chimique continu afin de construire un modèle pour la phase d’apprentissage de méthode et localisation et détection de pannes. Un dispositif expérimental a été conçu pour disposer de données expérimentales significatives. Pour le diagnostique et la localisation des méthodes orientées données ont été retenues, principalement les réseaux Bayésiens et les réseaux de neurones à Fonctions Radiales de Base (RBF) couplés à un algorithme génétique auto adaptatif à ajustement local (GAAPA). Les données collectées à partir du dispositif expérimental ont servi à l’apprentissage et à la validation du modèle. / The aim is to study a continuous chemical process, and then analyze the hold process of the reactor and build the models which could be trained to realize the fault diagnosis and localization in the process. An experimental system has been built to be the research base. That includes experiment part and record system. To the diagnosis and localization methods, the work presented the methods with the data-based approach, mainly the Bayesian network and RBF network based on GAAPA (Genetic Algorithm with Auto-adapted of Partial Adjustment). The data collected from the experimental system are used to train and test the models.

La simulation numérique est couramment utilisée pour étudier le comportement d’un composant et optimiser sa conception. Pour autant, chaque calcul est souvent coûteux en termes de temps et l’optimisation nécessite de résoudre un grand nombre de fois le modèle numérique pour différentes configurations du composant. Une solution actuelle pour réduire le temps de calcul consiste à remplacer la simulation coûteuse par un métamodèle. Des stratégies sont ensuite mises en place pour réaliser l’optimisation du composant à partir du métamodèle. Dans le cadre de cette thèse, trois dispositifs représentatifs des applications pouvant être traitées au sein du CEA LETI sont identifiés. L’étude de ces cas permet d’établir deux problématiques à résoudre. La première concerne la métamodélisation multi-fidélité, qui consiste à construire un métamodèle à partir de deux simulations du même composant ayant une précision différente. Les simulations sont obtenues à partir de différentes approximations du phénomène physique et aboutissent à un modèle appelé haute-fidélité (précis et coûteux) et un modèle basse fidélité (grossier et rapide à évaluer). Le travail sur cette méthode pour le cas de la cellule photoacoustique a amené au développement d’un nouveau métamodèle multifidélité basé sur les fonctions à base radiale. La deuxième problématique concerne la prise en compte des incertitudes de fabrication dans la conception de dispositifs photoniques. L’optimisation des performances de composants en tenant compte des écarts observés entre la géométrie désirée et la géométrie obtenue en fabrication a nécessité le développement d’une méthode spécifique pour le cas du coupleur adiabatique. / Numerical simulation is widely employed in engineering to study the behavior of a device and optimize its design. Nevertheless, each computation is often time consuming and, during an optimization sequence, the simulation code is evaluated a large number of times. An interesting way to reduce the computational burden is to build a metamodel (or surrogate model) of the simulation code. Adaptive strategies are then set up for the optimization of the component using the metamodel prediction. In the context of this thesis, three representative devices are identified for applications that can be encountered within the CEA LETI optics and photonics department. The study of these cases resulted in two problems to be treated. The first one concerns multifidelity metamodeling, which consists of constructing a metamodel from two simulations of the same component that can be hierarchically ranked in accuracy. The simulations are obtained from different approximations of the physical phenomenon. The work on this method for the case of the photoacoustic cell has generated the development of a new multifidelity surrogate model based on radial basis function. The second problem relate to the consideration of manufacturing uncertainties in the design of photonic devices. Taking into account the differences observed between the desired geometry and the geometry obtained in manufacturing for the optimization of the component efficiency requires the development of a particular method for the case of the adiabatic coupler. The entire work of this thesis is capitalized in a software toolbox.

This thesis presents the research and development of robust metamodelling tools for engineering design. Metamodelling in engineering is typically used for reducing computational cost of highly expensive analyses or simulations. Metamodels have been shown to be effective in these problems where an approximation constructed from a limited set of true data points is used in support of optimisation. The inspiration for this work is drawn from the optimisation of aircraft wing structures, constructed using large numbers of rectangular stiffened panels. When optimising such structures to produce a minimum weight design, it is necessary to evaluate multiple design constraints such as buckling load, damage tolerance and repairability. The total computational cost for this aspect of the analysis can become considerable when a large number of evaluations is required and can creates a bottleneck in the optimisation workflow. In response to this industrial design problem, a specification is proposed for an efficient and adaptive metamodelling formulation. Following an extensive literature review the multilevel Radial Basis Function (mRBF) model is highlighted as a promising candidate for further investigation. The mRBF formulation is discussed in detail, and a comparative study is presented comparing mRBF to more established modelling techniques. mRBF is then put to work on a range of optimisation test problems, including an industrial scale multi-panel wing design scenario. Emphasis is placed on the adaptive acquisition of model data as the optimisation process progresses. Implementation details and software development processes are also presented in detail. The case is made for decoupled modelling workflows, and a RESTful web based mRBF modelling framework. Finally the performance of the proposed modelling scheme is compared to the original specification, and recommendations are made for further investigation.

Drug development is a long and expensive process. It is often not until potential drug candidates are administered to humans that accurate quantification of their pharmacokinetic characteristics is achieved. The goal of developing quantitative structure-pharmacokinetic relationships (QSPkRs) is to relate the molecular structure of a chemical entity with its pharmacokinetic characteristics. In this thesis artificial neural networks (ANNs) were used to construct in silico predictive QSPkRs for various pharmacokinetic parameters using different drug data sets. Drug pharmacokinetic data for all studies were taken from the literature. Information for model construction was extracted from drug molecular structure. Numerous theoretical descriptors were generated from drug structure ranging from simple constitutional and functional group counts to complex 3D quantum chemical numbers. Subsets of descriptors were selected which best modeled the target pharmacokinetic parameter(s). Using manual selective pruning, QSPkRs for physiological clearances, volumes of distribution, and fraction bound to plasma proteins were developed for a series of beta-adrenoceptor antagonists. All optimum ANN models had training and cross-validation correlations close to unity, while testing was performed with an independent set of compounds. In most cases the ANN models developed performed better than other published ANN models for the same drug data set. The ability of ANNs to develop QSPkRs with multiple target outputs was investigated for a series of cephalosporins. Multilayer perceptron ANN models were constructed for prediction of half life, volume of distribution, clearances (whole body and renal), fraction excreted in the urine, and fraction bound to plasma proteins. The optimum model was well able to differentiate compounds in a qualitative manner while quantitative predictions were mostly in agreement with observed literature values. The ability to make simultaneous predictions of important pharmacokinetic properties of a compound made this a valuable model. A radial-basis function ANN was employed to construct a quantitative structure-bioavailability relationship for a large, structurally diverse series of compounds. The optimum model contained descriptors encoding constitutional through to conformation dependent solubility characteristics. Prediction of bioavailability for the independent testing set were generally close to observed values. Furthermore, the optimum model provided a good qualitative tool for differentiating between drugs with either low or high experimental bioavailability. QSPkR models constructed with ANNs were compared with multilinear regression models. ANN models were shown to be more effective at selecting a suitable subset of descriptors to model a given pharmacokinetic parameter. They also gave more accurate predictions than multilinear regression equations. This thesis presents work which supports the use of ANNs in pharmacokinetic modeling. Successful QSPkRs were constructed using different combinations of theoretically-derived descriptors and model optimisation techniques. The results demonstrate that ANNs provide a valuable modeling tool that may be useful in drug discovery and development.

A meshfree approximation scheme based on the radial basis function methods is presented for the numerical solution of the options pricing model. This thesis deals with the valuation of the European, Barrier, Asian, American options of a single asset and American options of multi assets. The option prices are modeled by the Black-Scholes equation. The θ-method is used to discretize the equation with respect to time. By the next step, the option price is approximated in space with radial basis functions (RBF) with unknown parameters, in particular, we con- sider multiquadric radial basis functions (MQ-RBF). In case of Ameri- can options a penalty method is used, i.e. removing the free boundary is achieved by adding a small and continuous penalty term to the Black- Scholes equation. Finally, a comparison of analytical and finite difference solutions and numerical results from the literature is included.

Radial basis function (RBF) approximation, is a new extremely powerful tool that is promising for high-dimensional problems, such as those arising from pricing of basket options using the Black-Scholes partial differential equation. The main problem for RBF methods have been ill-conditioning as the RBF shape parameter becomes small, corresponding to flat RBFs. This thesis employs a recently developed method called the RBF-QR method to reduce computational cost by improving the conditioning, thereby allowing for the use of a wider range of shape parameter values. Numerical experiments for the one-dimensional case are presented  and a MATLAB implementation is provided. In our thesis, the RBF-QR method performs better  than the RBF-Direct method for small shape parameters. Using Chebyshev points, instead of a standard uniform distribution, can increase the accuracy through clustering of the nodes towards the boundary. The least squares formulation for RBF methods is preferable to the collocation approach because it can result in smaller errors  for the same number of basis functions.

In this thesis, a popular music genre recognition approach for Japanese popular music using SVM (support vector machine) with forward feature selection is proposed. First, various common acoustic features are extracted from the digital signal of popular music songs, including sub-bands, energy, rhythm, tempo, formants. A set of the most appropriate features for the genre identification is then selected by the proposed forward feature selection technique. Experiments conducted on the database consisting of 296 Japanese popular music songs demonstrate that the accuracy of recognition the proposed algorithm can achieve approximately 78.81% and the accuracy is stable when the number of testing music songs is increased.

This dissertation introduces novel methods for solving highly challenging model- ing and control problems, motivated by advanced aerospace systems. Adaptable, ro- bust and computationally effcient, multi-resolution approximation algorithms based on Radial Basis Function Network and Global-Local Orthogonal Mapping approaches are developed to address various problems associated with the design of large scale dynamical systems. The main feature of the Radial Basis Function Network approach is the unique direction dependent scaling and rotation of the radial basis function via a novel Directed Connectivity Graph approach. The learning of shaping and rota- tion parameters for the Radial Basis Functions led to a broadly useful approximation approach that leads to global approximations capable of good local approximation for many moderate dimensioned applications. However, even with these refinements, many applications with many high frequency local input/output variations and a high dimensional input space remain a challenge and motivate us to investigate an entirely new approach. The Global-Local Orthogonal Mapping method is based upon a novel averaging process that allows construction of a piecewise continuous global family of local least-squares approximations, while retaining the freedom to vary in a general way the resolution (e.g., degrees of freedom) of the local approximations. These approximation methodologies are compatible with a wide variety of disciplines such as continuous function approximation, dynamic system modeling, nonlinear sig-nal processing and time series prediction. Further, related methods are developed for the modeling of dynamical systems nominally described by nonlinear differential equations and to solve for static and dynamic response of Distributed Parameter Sys- tems in an effcient manner. Finally, a hierarchical control allocation algorithm is presented to solve the control allocation problem for highly over-actuated systems that might arise with the development of embedded systems. The control allocation algorithm makes use of the concept of distribution functions to keep in check the "curse of dimensionality". The studies in the dissertation focus on demonstrating, through analysis, simulation, and design, the applicability and feasibility of these ap- proximation algorithms to a variety of examples. The results from these studies are of direct utility in addressing the "curse of dimensionality" and frequent redundancy of neural network approximation.

Evolution of neural networks, or neuroevolution, has been a successful approach to many low-level control problems such as pole balancing, vehicle control, and collision warning. However, certain types of problems — such as those involving strategic decision-making — have remained difficult to solve. This dissertation proposes the hypothesis that such problems are difficult because they are fractured: The correct action varies discontinuously as the agent moves from state to state. To evaluate this hypothesis, a method for measuring fracture using the concept of function variation of optimal policies is proposed. This metric is used to evaluate a popular neuroevolution algorithm, NEAT, empirically on a set of fractured problems. The results show that (1) NEAT does not usually perform well on such problems, and (2) the reason is that NEAT does not usually generate local decision regions, which would be useful in constructing a fractured decision boundary. To address this issue, two neuroevolution algorithms that model local decision regions are proposed: RBF-NEAT, which biases structural search by adding basis-function nodes, and Cascade-NEAT, which constrains structural search by constructing cascaded topologies. These algorithms are compared to NEAT on a set of fractured problems, demonstrating that this approach can improve performance significantly. A meta-level algorithm, SNAP-NEAT, is then developed to combine the strengths of NEAT, RBF-NEAT, and Cascade-NEAT. An evaluation in a set of benchmark problems shows that it is possible to achieve good performance even when it is not known a priori whether a problem is fractured or not. A final empirical comparison of these methods demonstrates that they can scale up to real-world tasks like keepaway and half-field soccer. These results shed new light on why constructive neuroevolution algorithms have difficulty in certain domains and illustrate how bias and constraint can be used to improve performance. Thus, this dissertation shows how neuroevolution can be scaled up from learning low-level control to learning strategic decision-making problems. / text

Drug development is a long and expensive process. It is often not until potential drug candidates are administered to humans that accurate quantification of their pharmacokinetic characteristics is achieved. The goal of developing quantitative structure-pharmacokinetic relationships (QSPkRs) is to relate the molecular structure of a chemical entity with its pharmacokinetic characteristics. In this thesis artificial neural networks (ANNs) were used to construct in silico predictive QSPkRs for various pharmacokinetic parameters using different drug data sets. Drug pharmacokinetic data for all studies were taken from the literature. Information for model construction was extracted from drug molecular structure. Numerous theoretical descriptors were generated from drug structure ranging from simple constitutional and functional group counts to complex 3D quantum chemical numbers. Subsets of descriptors were selected which best modeled the target pharmacokinetic parameter(s). Using manual selective pruning, QSPkRs for physiological clearances, volumes of distribution, and fraction bound to plasma proteins were developed for a series of beta-adrenoceptor antagonists. All optimum ANN models had training and cross-validation correlations close to unity, while testing was performed with an independent set of compounds. In most cases the ANN models developed performed better than other published ANN models for the same drug data set. The ability of ANNs to develop QSPkRs with multiple target outputs was investigated for a series of cephalosporins. Multilayer perceptron ANN models were constructed for prediction of half life, volume of distribution, clearances (whole body and renal), fraction excreted in the urine, and fraction bound to plasma proteins. The optimum model was well able to differentiate compounds in a qualitative manner while quantitative predictions were mostly in agreement with observed literature values. The ability to make simultaneous predictions of important pharmacokinetic properties of a compound made this a valuable model. A radial-basis function ANN was employed to construct a quantitative structure-bioavailability relationship for a large, structurally diverse series of compounds. The optimum model contained descriptors encoding constitutional through to conformation dependent solubility characteristics. Prediction of bioavailability for the independent testing set were generally close to observed values. Furthermore, the optimum model provided a good qualitative tool for differentiating between drugs with either low or high experimental bioavailability. QSPkR models constructed with ANNs were compared with multilinear regression models. ANN models were shown to be more effective at selecting a suitable subset of descriptors to model a given pharmacokinetic parameter. They also gave more accurate predictions than multilinear regression equations. This thesis presents work which supports the use of ANNs in pharmacokinetic modeling. Successful QSPkRs were constructed using different combinations of theoretically-derived descriptors and model optimisation techniques. The results demonstrate that ANNs provide a valuable modeling tool that may be useful in drug discovery and development.