Evolutionary tuning of neural networks for gesture recognition

This paper is about a data glove/neural network system as a powerful input device for virtual reality and multi media applications. In contrast to conventional keyboards, space balls, and two-dimensional mice, which allow for only rudimental inputs, the data glove system allows the user to present the system with a rich set of intuitive commands. Previous research has employed different neural networks to recognize various hand gestures. Due to their on-line adaptation capabilities, radial basis function networks are preferably over backpropagation. Unfortunately, the latter have shown better recognition rates. This paper applies evolutionary algorithms to fine tune pre-learned radial basis function networks. After optimization, the networks achieves a recognition rate of up to 100%, and is therefore comparable or even better than that of backpropagation networks.