Search

International audience; This study aims to aid understanding of Model Predictive Control (MPC) alternatives through comparing most interesting MPC implementations. This comparison will be performed intrinsically and illustrated using the four-tank benchmark, widely studied by academics taking care of industrial perspectives. Although MPC provides advanced control solutions for a wide class of dynamical systems, challenges arise in managing the compromise between accuracy, computational cost and resilience, depending on the type of model used. In this study, linear, linear time-varying and non-linear MPCs are compared to MPC that uses a neural network based predictive model identified from data. The tuning and implementation methods considered are discussed, and accurate simulation results provided and analyzed. Precisely, the performance of each method (linear, linear time-varying, non-linear MPC) are compared to the neural MPC. Pros and cons of neural MPC are highlighted.

International audience; This paper deals with model predictive control synthesis which take benefits from artificialneural networks to model (non-linear) dynamical system. More precisely, thanks to a systematic andrigorous methodology, it is shown that residual networks (ResNet) and PolyInception networks (PolyNet)neural network architectures, developed initially for image recognition, are very good candidate for i)identification of dynamical systems, ii) being used as embedded model in a model predictive control laws.Concretely, the widely used non-linear dynamical system quadruple tank process is used as a benchmark.The neural network architectures studied are i) feedforward networks as a reference point, and the two otherlinked to Euler integration method ii) residual networks and iii) PolyInception networks. Networks trainingis performed by mixing classical back-propagation algorithm and hyperparameters optimisation throughheuristics. The identification results provided show that neural networks of types ii) and iii) perform betterthan the classical one i), with a better generalisation capability. Finally, model predictive controllers aresynthesized based on the various networks trained. The simulation results obtained for controlling waterlevels of a 4 tanks system benchmark give interesting insights. They show that residual networks basedmodel predictive control is better suited than feedforward networks and PolyInception networks based ones,both taking into account computation time and set point errors.

Simulating spiking neuron networks with a sequential event-driven approach consumes less computation time than clock-driven methods. On the other hand, a parallel computing support provides a larger amount of material ressources for optimizing simulation performance. This PhD dissertation proposes an event-driven, multithreaded and distributed framework for simulating large size spiking neuron networks. The name of the simulator is the acronym DAMNED, for Distributed And Multithreaded Neural Event-Driven simulation framework. DAMNED distributes the neurons and connections of the network on the material ressources synchronized through a decentralized global virtual time. DAMNED also couples local multithreaded processing to the distributed hardware. DAMNED allows to speed up the simulation and to manage wider neural networks than sequential processing. DAMNED is suited to run many models of spiking neurons and networks, and most material supports are workable. Using DAMNED is presented first on simple networks for different sizes, connectivities and activities. Next, DAMNED is applied to modelling the control of saccadic eye movements. Completely based on spiking neurons, the model studies interactions between the neural circuits of the saccadic system located in the brainstem. The model helps validating the hypothesis that the saccade amplitude could be encoded by a vector summation of the activities in the superior colliculus motor map rather than a vector average, from comparison with data obtained in the simulation. Even if further developments and improvements may be forecasted, the originality of the work is to couple event-driven and distributed programming. Moreover, among the parallel simulators for spiking neuron networks, DAMNED is the first one taking advantage of an event-driven strategy internal multithreading of the logic processes and a distributed architecture of physical processes. Hence DAMNED is an advance in the area of simulating spiking neuron networks, mainly for wide size networks. Experiments on simulating the control of saccadic eye movements by a spiking neural network model and their contributions for the neuroscience community confirm the perspectives for further uses of the present work; Ce travail de thèse propose un simulateur événementiel, multithreadé et distribué pour la simulation de réseaux de neurone impulsionnels de grande taille, nommé DAMNED, qui signifie Distributed And Multithreaded Neural Event-Driven simulation framework. Répartissant le réseau de neurones sur les ressources matérielles synchronisées par une méthode décentralisée de gestion du temps virtuel, DAMNED introduit également un fonctionnement multithread. DAMNED permet d'accélérer les calculs et de simuler des réseau de plus grande taille qu'en séquentiel. DAMNED pennet d'exploiter de nombreux modèles de réseaux et de neurones et la plupart des supports matériels sont exploitables. Nous présentons l'utilisation de DAMNED sur un modèle simple de réseau pour différentes taille, connectivités et dynamiques. Ensuite, nous proposons une application directe de DAMNED dans une modélisation du système saccadique restreint au tronc cérébral. On montre, à l'aide de ce modèle, que l'hypothèse selon laquelle une somme vectorielle (vectc summation) des activités de la carte motrice du colliculus supérieur coderait pour l'amplitude de la saccade correspond davantage au données obtenues pour le modèle, exécuté sur DAMNED, qu'à l'hypothèse d'un moyennage de vecteurs (vector average). L'originalité de ce travail, parmi les premiers simulateurs distribués de réseaux de neurones impulsionnels, réside dans le couplage d'une stratégie événementielle, d'un multithreading interne aux processus logiques et une architecture physique distribuée. Le simulateur DAMNE D constitue donc une avancée dans le domaine des réseaux de neurones impulsionnels de grande taille

Simulating spiking neuron networks with a sequential event-driven approach consumes less computation time than clock-driven methods. On the other hand, a parallel computing support provides a larger amount of material ressources for optimizing simulation performance. This PhD dissertation proposes an event-driven, multithreaded and distributed framework for simulating large size spiking neuron networks. The name of the simulator is the acronym DAMNED, for Distributed And Multithreaded Neural Event-Driven simulation framework. DAMNED distributes the neurons and connections of the network on the material ressources synchronized through a decentralized global virtual time. DAMNED also couples local multithreaded processing to the distributed hardware. DAMNED allows to speed up the simulation and to manage wider neural networks than sequential processing. DAMNED is suited to run many models of spiking neurons and networks, and most material supports are workable. Using DAMNED is presented first on simple networks for different sizes, connectivities and activities. Next, DAMNED is applied to modelling the control of saccadic eye movements. Completely based on spiking neurons, the model studies interactions between the neural circuits of the saccadic system located in the brainstem. The model helps validating the hypothesis that the saccade amplitude could be encoded by a vector summation of the activities in the superior colliculus motor map rather than a vector average, from comparison with data obtained in the simulation. Even if further developments and improvements may be forecasted, the originality of the work is to couple event-driven and distributed programming. Moreover, among the parallel simulators for spiking neuron networks, DAMNED is the first one taking advantage of an event-driven strategy internal multithreading of the logic processes and a distributed architecture of physical processes. Hence DAMNED is an advance in the area of simulating spiking neuron networks, mainly for wide size networks. Experiments on simulating the control of saccadic eye movements by a spiking neural network model and their contributions for the neuroscience community confirm the perspectives for further uses of the present work, Ce travail de thèse propose un simulateur événementiel, multithreadé et distribué pour la simulation de réseaux de neurone impulsionnels de grande taille, nommé DAMNED, qui signifie Distributed And Multithreaded Neural Event-Driven simulation framework. Répartissant le réseau de neurones sur les ressources matérielles synchronisées par une méthode décentralisée de gestion du temps virtuel, DAMNED introduit également un fonctionnement multithread. DAMNED permet d'accélérer les calculs et de simuler des réseau de plus grande taille qu'en séquentiel. DAMNED pennet d'exploiter de nombreux modèles de réseaux et de neurones et la plupart des supports matériels sont exploitables. Nous présentons l'utilisation de DAMNED sur un modèle simple de réseau pour différentes taille, connectivités et dynamiques. Ensuite, nous proposons une application directe de DAMNED dans une modélisation du système saccadique restreint au tronc cérébral. On montre, à l'aide de ce modèle, que l'hypothèse selon laquelle une somme vectorielle (vectc summation) des activités de la carte motrice du colliculus supérieur coderait pour l'amplitude de la saccade correspond davantage au données obtenues pour le modèle, exécuté sur DAMNED, qu'à l'hypothèse d'un moyennage de vecteurs (vector average). L'originalité de ce travail, parmi les premiers simulateurs distribués de réseaux de neurones impulsionnels, réside dans le couplage d'une stratégie événementielle, d'un multithreading interne aux processus logiques et une architecture physique distribuée. Le simulateur DAMNE D constitue donc une avancée dans le domaine des réseaux de neurones impulsionnels de grande taille

National audience; Cet article présente le simulateur DAMNED, destiné à la simulation rapide de grands réseaux de neurones impulsionnels.Les choix qui sont faits, sur le plan computationnel, ont pour objectif un fonctionnement efficace, sur machine parallèle, réelle ou simulée. La première partie expose les grands principes de fonctionnement du système : processus distribués et multithreadés, horloge virtuelle, gérée par échanges de messages-événements. La question de la mise en oeuvre de divers modèles de neurones et de réseaux est ensuite abordée. Des mesures de temps sont présentées, sur une station de travail et sur un parc de PC, pour des réseaux de plusieurs milliers de neurones dont les poids sont modifiés par STDP (Spike-Time-Dependent Plasticity).

In a Spiking Neural Networks (SNN), spike emissions are sparsely and irregularly distributed both in time and in the network architecture. Since a current feature of SNNs is a low average activity, efficient implementations of SNNs are usually based on an Event-Driven Simulation (EDS). On the other hand, simulations of large scale neural networks can take advantage of distributing the neurons on a set of processors (either workstation cluster or parallel computer). This article presents DAMNED, a large scale SNN simulation framework able to gather the benefits of EDS and parallel computing. Two levels of parallelism are combined: Distributed mapping of the neural topology, at the network level, and local multithreaded allocation of resources for simultaneous processing of events, at the neuron level. Based on the causality of events, a distributed solution is proposed for solving the complex problem of scheduling without synchronization barrier., Comment: 6 pages