Reducción de Tiempos de Entrenamiento de Algoritmos de Aprendizaje de Máquina a través de Tutoría por parte de un Experto Externo

Abstract

This degree project presents the approach and implementation of a learning policy for neural networks, based on and refining existing techniques trained through the experience of an external agent (expert), taking as reference a convolutional neural network architecture. The selection process of the problematic environment, the architecture and the selected paradigm are detailed. Likewise, multiple tests are performed to confirm the performance. They are trained with different training policies. Finally, the performance of the different trained policies is evaluated with respect to a base policy \textit{exploration/exploitation}. As a final product we present: the reference expert dataset, a repository with the programs made for the development together with the implementation of the policy.