Neural Network Toolbox

Thoughts and Conclusions

While newlind is not able to return a zero error solution for nonlinear problems, it does minimize the sum-squared error. In many cases, the solution, while not perfect, may model a nonlinear relationship well enough to meet the application specifications. Giving the linear network many delayed signal values gives it more information with which to find the lowest error linear fit for a nonlinear problem.

Of course, if the problem is very nonlinear and/or the desired error is very low, backpropagation or radial basis networks would be more appropriate.

Network Testing

Applin2: Adaptive Prediction

© 1994-2005 The MathWorks, Inc.