Recurrent Flexible Neural Tree Model for Time Series Prediction.

In this paper, a new encoding schemes based on tree representation is represented to encode recurrent multi layer neural network. It implement a learning process formed by two iterative phases: structure optimization and parameters optimization. For the structure evolving, a modified version of the Genetic Programming algorithm was adapted to support the recurrent topology of the network. On the other hand, a hybrid version of Harmony Search algorithm is used to adjust the network parameters including connection weights and neurons parameter set. Besides, the proposed model is evaluated by dynamical chaotic times series and compared with other studies.