next up previous contents
Next: Global Optimization Up: Use of Second Previous: Use of Second

Memory Constraints

Implementations of the LM algorithm and the Quasi-Newton algorithms require intermediate storage of n x n matrices, compared to only a n element vector for the Conjugate Gradient Descent algorithm. When n is very large (such as for neural networks), this can become an issue and the latter algorithm may be preferred.

Bob Fisher
Fri Mar 28 14:12:50 GMT 1997