Aplicação de evolução estrutural e paramétrica de redes neurais dinâmicas na melhoria de desempenho de métodos de assimilação de dados

Abstract

The use of numerical prediction models are essential to modern society. Data assimilation is a technique that aims to increase the efficiency of prediction of these models by combining model data with data from observations, obtaining a state that is closer to the true state of nature. Combining these two sources of information, observational and model data, has been a challenge, even for supercomputers present in this type of application. Thus neural networks have been proposed as an alternative to perform the assimilation of high quality at a lower computational cost. This paper proposes to investigate the neuroevolucionist model NEAT in data assimilation. NEAT is able to adapt, using principles of evolutionary computation, the weights of the connections and the topology of the neural network in a search for a minimum topology and getting better performance. So, it was developed a software that enabled the test and evaluation of the proposed approach. Through the experiment on the Lorenz Attractor was found that the model NEAT was able to emulate the task of data assimilation with a smaller error when compared with the neural network trained by backpropagation from the experiment on the Shallow Water model was observed that NEAT model always gets a topology with significantly fewer operations and states with large enough also operates at a lower computational cost.