Back-propagation with chaos

Multilayer feed-forward neural networks are widely used based on minimization of an error function. Back-propagation is a famous training method used in the multilayer networks but it often suffers from a local minima problem. To avoid this problem, we propose a new back-propagation training based on chaos. We investigate whether randomicity and ergodicity property of chaos can enable the learning algorithm to escape from local minima. Validity of the proposed method is examined by performing simulations on three real classification tasks, namely, the Ionosphere, the Wincson Breast Cancer (WBC), and the credit-screening datasets. The algorithm is shown to work better than the original back-propagation and is comparable with the Levenberg-Marquardt algorithm, but simpler and easier to implement comparing to Levenberg-Marquardt algorithm.