Training of Recurrent ANN with Time Decreased Activation by GA to the Forecast in Dynamic Problems

摘要

In this paper, we state an evolution of the recurrent ANN (RANN) to enforce the persistence of activations within the neurons to create activation contexts that generate correct outputs through time. In this new focus we want to file more information in the neuron's connections. To do this, the connection's representation goes from the unique values up to a function that generates the neuron's output. The training process to this type of ANN has to calculate the gradient that identifies the Junction. To train this RANN we developed a GA based system that find the best gradient set to solve each problem.