Second-order recurrent neural network for word sequence learning

This paper presents a genetic algorithm (GA)-based 2^ndorder recurrent neural network (GRNN). Feedbacks in the structure enable the network to remember cues from the recent past of a word sequence. The GA is used to help design an improved network by evolving weights and connections dynamically. Simulation results on learning 50 commands of up to 3 words and 24 phone numbers of 10 digits illustrate that the GRNN is most efficient in error performance and recall accuracy as compared to other backpropagation-based recurrent and feedforward networks. The effects of population size, crossover probability and mutation rate on the performance of the GRNN are presented