One of the biggest problems I have had with all Nnet platforms is that I'll make a run where an input parameter is not what I thought it was and I get a superior result by accident. In most cases, I don't know it at the time and by the time I do realize there was something "miss-set" it is long after the fact and there is no longer any way to determine how the run was setup.
With Finite Element Programs, it is customary to include a copy of the input file(s) in the output so that you know exactly what input caused what output.
As it stands now, there is no way to actually prove what a Membrain input was unless you simply redo the analysis. While this sounds good on the surface, I am not completely sure, but I think you would not get the same initial weight randomization which means you cannot redo the exact same analysis. This makes it exceeding difficult when you have a model that is sensitive to initial weights.
NOTE: The best randomization method for repeatability is to use the same random sequence for each of a set of random number seeds. For example, the random series for seed #57 will always be the same. If you have 4 weights, the first 4 terms would be used for initialization - if you have 250 weights, the first 250 terms would be used, and so on. This would greatly facilitate the development of models of difficult systems where repeatability is necessary to evaluate model changes.
Thanks
Tom
An Input Settings Run Log File and Random Number Seeds
-
- Posts: 55
- Joined: Wed 21. Jul 2010, 18:43