The University of Sussex

A Marquardt Algorithm for choosing the step-size in backpropagation learning with conjugate gradients

Peter M. Williams

Standard learning in feed-forward networks uses simple gradient descent, sometimes with a "momentum" term. Gradient descent is very inefficient. The momentum method is an improvement though it is ad hoc and shares with the steepest descent method the disadvantage of requiring an arbitrary choice of parameter.


Download compressed postscript file