Hi Mr. Heaton,
I've found the following thread about continuous training in Encog 2.x:
I've a similar question for Encog 3.1 – I'd like to ask you whether the mentioned problem (lost deltas on exit and restart) is already solved and whether there is some level of support for continuous training (e.g. persistence of Train classes or anything else)?
Is it possible to continue training a network loaded from an EG file?
I have the following idea:
I want to perform temporal predictions for multiple data sources (temperature predictions for multiple temperature probes). New values from data sources comes continuously and future values are predicted at the same time – it should be possible according to linked forum thread. When I add a new data source, I'd like to use an existing NN for it. There would be multiple application instances (for load balancing) so NN for new data sources will be interchanged between them probably via EG files.
Problem is that waveforms (data characteristics) can slightly differ depending to particular data source so a NN used for the new data source has to be retrained. But I have no data set available for training yet and I don't want to wait for gathering a sufficiently large data set for training. So I'd like to provide new values as they comes in time from the data source and use them for continuous training of NN making its' predictions more and more accurate.
Is this possible with Encog 3.1, or are there some hitches like lost deltas on exit and restart?
Thanks for your answer and also for research and development of Encog!
d like to provide new values as they comes in time from the data source and use them for continuous training of NN making its