Nuevo Modelo de Red Neuronal para Aprendizaje Supervisado Basado en Aprendizaje por Refuerzo con Valores de Influencia

• Autor: Valdivia Ballesteros, Andre´ Mauricio
• Formato: tesis de maestría
• Fecha de Publicación: 2018
• Institución: Consejo Nacional de Ciencia Tecnología e Innovación
• Repositorio: CONCYTEC-Institucional
• Lenguaje: español
• OAI Identifier: oai:repositorio.concytec.gob.pe:20.500.12390/1678
• Enlace del recurso: https://hdl.handle.net/20.500.12390/1678
• Nivel de acceso: acceso abierto
• Materia: Redes neuronales; Multiagentes; Aprendizaje por Refuerzo; Auto- organización; https://purl.org/pe-repo/ocde/ford#2.02.03

Neural self-organization is an innate feature of the brains of mammals and it isvery necessary for their operation. The most known artificial neural network models that use this characteristic are the Self-Organized Maps (SOM) and the Adaptive Resonance Theory (ART), but these models do not take the neuron as a processing unit, as it’s biological counterpart does; besides these are models mostlyused for the unsupervised learning paradigm, this means that there aren’t robustself-organized models in the supervised learning paradigm. In other way, influence value reinforcement learning paradigm, used in multi-agent systems, prove thatagents can communicate among them, and can self-organize themselves to assigntasks, without interference.Motivated by the lack of features in the artificial neural networks, and taking intoaccount the influence values reinforcement algorithm, a new neural network modelis proposed, which is focused on solving supervised learning problems by usingreinforcement learning agents as neurons in our model; model that has the differentactivation functions as an important characteristic, because these are unique foreach neuron. This is also an important feature for self-organization.The neural agents will work in a discrete space, besides using a learning algorithm different from the backpropagation, which is used in many networks. Analgorithm inspired in the way the SOM networks propagate their knowledge isproposed, this way the neighboring states to the trained state can acquire its knowledge.In order to prove the functionality of this model, low dimensionality daya bases were used and their performance was compared by a multilayer perceptron,where in most of the databases its performance was improved. The creation of thisnew model is the base for further importance of this investigation is the concept ofneuron.To prove the model functionality, we used databases of low dimensionality, andwe compare its performance with multilayer perceptron, where in most of the databases the performance was improved. The creation of this novel model, is the basefor further research, where the fundamental importance of this work is the novelneural concept