Search

International audience; The use of neural networks in finance began by the end of the 1980s and by the beginning of the 1990s, it developed specific applications related to forecasting on the failure of companies. In order to highlight the evolution of this research stream, we have retained and analysed 30 studies in which the authors use neural networks to solve companies' classification problems (healthy and failing firms). This review of all these works gives us the opportunity to stress upon future trends in bankruptcy forecasting research

This thesis studies the use of connectionist algorithms for solving reinforcement learning problems. Connectionist algorithms are inspired by the way information is processed by the brain: they rely on a large network of highly interconnected simple units, which process numerical information in a distributed and massively parallel way. Reinforcement learning is a computational theory that describes the interaction between an agent and an environment: it enables to precisely formalize goal-directed learning from interaction.We have considered three problems, with increasing complexity, and shown that they can be solved with connectionist algorithms: 1) Reinforcement learning in a small state space: we exploit a well-known algorithm in order to build a connectionist network: the problem's paramaters are stored into weighted units and connections and the planning is the result of a distributed activity in the network. 2) Learning a representation for approximating a reinforcement learning problem with a large state space: we provide an algorithm for automatically building a state space partition in order to approximate a large problem. 3) Self-organization of specialized modules for approximating various reinforcement problems with a large state space: we exploit a ``divide and conquer'' approach and show that various tasks can efficiently be spread over a little number of specialized functional modules.; Cette thèse étudie l'utilisation d'algorithmes connexionnistes pour résoudre des problèmes d'apprentissage par renforcement. Les algorithmes connexionnistes sont inspirés de la manière dont le cerveau traite l'information : ils impliquent un grand nombre d'unités simples fortement interconnectées, manipulant des informations numériques de manière distribuée et massivement parallèle. L'apprentissage par renforcement est une théorie computationnelle qui permet de décrire l'interaction entre un agent et un environnement : elle permet de formaliser précisément le problème consistant à atteindre un certain nombre de buts via l'interaction.Nous avons considéré trois problèmes de complexité croissante et montré qu'ils admettaient des solutions algorithmiques connexionnistes : 1) L'apprentissage par renforcement dans un petit espace d'états : nous nous appuyons sur un algorithme de la littérature pour construire un réseau connexionniste ; les paramètres du problème sont stockés par les poids des unités et des connexions et le calcul du plan est le résultat d'une activité distribuée dans le réseau. 2) L'apprentissage d'une représentation pour approximer un problème d'apprentissage par renforcement ayant un grand espace d'états : nous automatisons le procédé consistant à construire une partition de l'espace d'états pour approximer un problème de grande taille. 3) L'auto-organisation en modules spécialisés pour approximer plusieurs problèmes d'apprentissage par renforcement ayant un grand espace d'états : nous proposons d'exploiter le principe "diviser pour régner" et montrons comment plusieurs tâches peuvent être réparties efficacement sur un petit nombre de modules fonctionnels spécialisés.

International audience; The design and topology of a neural network is still an important and difficult task. To solve the problems of topology posed by the introduction of connexionism, new approaches are proposed, and especially a combination of induction rules with a statistical estimation of the neuron coefficients for each layer. This research aims to compare an algorithm of this SLN approach with traditional methods (regression and classical BP neural networks) using the gradient method. Methods are put into application to determine the price-quality relationship of a complex product, the automobile, according to the hedonic price model. This application of the price-quality relationship to the English automobile market leads to the conclusion that the claimed superiority of this approach is unsubstantiated since, compared to the BP neural networks and even linear regression, the performance of the GMDH method is inferior.

Contribution à un ouvrage.; Whereas classical connectionist models can hardly cope with difficult dynamic tasks with a strong temporal factor, many temporal mechanisms inspired with neurobiological data has been proposed in the past and yield efficient time processing properties. The goal of this chapter is to show that, beyond these isolated mechanisms, their integration in a more general architectural and functional framework can potentiate their power and make them usable for non trivial behavioral tasks. We propose a cerebral framework, from the neuronal to the behavioral level, and give some applicative illustrations that underline the encouraging results obtained today.