A Recurrent Neural Network (RNN) is a neural network (NN) with recurrent connections. These connections forward part of the output signal back in network input. Since the NNN could stimulate its itself these neuron it could :

Such behaviors are complex and hard to control. Therefore a serie of custom RNN have been developped by the past : where RRN was componed of several NN bind together, or with a single recurrent connection... Another approch, used by NEAT project, is to let the network evolve by itself using Genetic Algorithm (GA). We choose this one.

JavaRNN use full connected RNNs. Such a RNN contains X nodes and X² connections. Each of this connection is weighted. The choosen transfer function is at this time a simple sigmoid.

Application

We actually use RNN as predictor : we check the RNN capability to predict a signal at T+1. We also use it as long-run predcitor, coupling its predicted output to its input we let im predict a signal at T+n

Delay

We have experienced limitations modelling periodic signals due to the natural resonant frequency of the RNN. To solve this problem we had delay buffers in each connection in form of a FIFO queue. the length of the queue is also set during the training.