Comparison of Nonlinear Optimisation Strategies for Feed-Forward Adaptive Layered Networks

摘要

Learning strategies which may be employed for the generic class of layered feed-forward adaptive networks exemplified by the traditional multilayer perceptron are discussed. Such a network is only useful if a set of weight values exists which allows the network to form a good approximation to an underlying (and possibly unknown) transformation between input and output patterns. The need for schemes capable of producing such a set of weights, (should one exist) as efficiently as possible is highlighted. The application of various learning algorithms to examples ranging from small scale trivial problems (solution of the XOR function), to larger scale pattern processing applications (speech recognition of isolated confusable whole words) is considered.