Dynamic memory by recurrent neural network and its learning by genetic algorithm

Recurrent neural networks have dynamic characteristics and can express functions of time. The recurrent neural networks can be applied to memorize robotic motions, i.e. trajectory of a manipulator. For this purpose, it is necessary to determine appropriate interconnection weights of the network. Formerly, learning algorithms based on gradient search techniques have been shown. However, it is difficult for the recurrent neural network to learn such functions while using previous approaches because of much computing requirement and limitation of memory. This paper presents a new learning scheme for the recurrent neural networks by genetic algorithm (GA). The GA is applied to determine interconnection weights of the recurrent neural networks. The GA approach is compared with the backpropagation through time which is a famous learning algorithm for the recurrent neural networks. Simulations illustrate the performance of the proposed approach