Using saving points in NN after additional training data has been added?

Hey guys,

I've got a neural network (NARX) to predict (tomorrow) prices of a product (one price for every hour of the day).

There are a lot of influeces, so the NN as a lot of inputs and right now there is no "best" NN architecture. This has been scientifically examined in a masters thesis, so right now there are multiple NN being calculated per day and the best of the day will be used.

Good thing is, the forecasts are pretty accurate. Bad thing is the calculation takes too long (~45 minutes per forecasted day).

The training data "grows" every day by one day (yesterdays data). This brought me to the idea, to use saving points (or something similiar) to not always have to train the NN from 0% to 100% and instead take the saving point, give the NN the new training data and let it start from there.

Is there any way to achieve this? Or will I still have to train the NN from 0 to 100% every day? Unfortunately, if I don't "search" for the best NN architecture every day, the fineness of the forecast decreases significantly.

Thank you very much in advance.

Answer 1

[Greg Heath]

I have had excellent success using RBFs that were continually updated over periods of years each time new data became available. The logic was somewhat similar to that of NEWRB. The most important exception was that the algorithm could be initialized with a specified set of basis functions. NEWRB can only start with an empty basis. However, the conversion should be relatively easy.

Hope this helps.

Thank you for formally accepting my answer

Greg