In SDR Classifier How long should we train? What will happen if we always train and predict ?
If we train and stop the training, will it be usable after some time? Thanks for help or guide in advance
1 Like
From my experiments I use the number of input buckets as criteria for stopping learning. If it is not changed for a given time interval, so we can stop it.
1 Like
Thanks for the reply.
Could you explain it ?
If the model is trained upto that criteria, will it effectively predict the future values for long?
For real application, for example predicting sinus wave, the input is float values at every time sample nT, where n = 1,2,… so that it may belong to a bucket or neighbor bucket index. Even though your sinus wave is periodical but you cannot get the same bucket sequence after many periods. The sequence length is changed significantly, so that you have to wait until all different bucket indices are found. Only after that TM can learn sequences and provide good results