>Have you considered using cursors? VFP cursors automatically span to disk, meaning you can forget about the resource problems to be expected from memory-resident collections, arrays or datasets. You can create thousands of VFP cursors if necessary- to store intermediate results or simplify nested processing. You can index your cursors on the fly after which you should be able to replace intensive loops with SQL queries that can be very quick against indexed tables or cursors.

Thank you for the suggestion. Some of these neural nets may have well over 5000 input nodes. Each set of weights is calcuated into a 5000 by 5000 array. Cursors and tables would be fine vertically, but as I recollect, horizontally their maximum is 256 fields which makes them unusable for this purpose. Standard backward propagation neural network code at this level will not execute in VFP.