An inverse model learning algorithm using the hierarchical mixtures of experts

A new learning algorithm in neural networks is proposed for inverse modeling. In the learning algorithm, the hierarchical mixtures of experts (HME) is tried out as a forward model of the system. The algorithm is fundamentally based on the back propagation procedure and the updating values of the network synaptic weights are calculated with the help of the HME. Almost all conventional learning algorithms use the Jacobian matrix of the system for estimating the neural network error. This is not the case with our algorithm. As a result, it carefully avoids the local minimum problem which often occurs in some inverse model learning processes