Multi-objective evolutionary Recurrent Neural Networks for system identification

This paper proposes a new multi-objective evolutionary approach for training recurrent neural networks (RNNs). The algorithm uses features of a variable length representation allowing easy adaptation of neural networks structures and a micro genetic algorithm (muGA) with an adaptive local search intensity scheme for local fine-tuning. In addition, a structural mutation (SM) operator for evolving the appropriate number of neurons for RNNs is used. Simulation results demonstrated the effectiveness of proposed method for system identification tasks.