A back propagation neural net capable of predicting crack depths in solid specimens using the results of NDE techniques is discussed. The focus of the research effort was the development of a faster convergence algorithm which would allow the neural net to be used for real time problems. Traditionally, the steepest descent method is used to solve the nonlinear multidimensional unconstrained optimization problem, but convergence is very slow. As a solution an adaptive step algorithm which varies the step size according to the topography of the error surface is developed. The problem of saturation is investigated and a solution found by modifying the weight update formulae for both the hidden and outer layers. When implemented, this in conjunction with the adaptive step algorithm, leads to vastly reduced convergence times. For the purposes of comparison, other methods using more formal optimization techniques e. g. Conjugate Gradient methods are implemented. The adaptive step algorithm is found to perform credit ably. Finally the robustness characteristics of the network are investigated, and certain interesting results obtained indicating that reaching a global minimum may not be the ideal solution in the context of pattern classification problems.

purchase the full-text of this paper (price £20)

go to the previous paper
go to the next paper
return to the table of contents
return to the book description
purchase this book (price £60 +P&P)