I recently spent some time in trying to figure out how to best use the SDRClassifier. The algorithm API example tells me to use activeCells as input to the classifier along with the current actual value to get 1-step ahead predictions. But then also the prediction of n-step ahead is generated by the SDRClassifier using these activeCells and a history of activeCells. I hope I am correct in this description.
If this is all true then the SDRClassifier is not just a classifier. Its is a predictor in and of itself. Calling it just a classifier is confusing. Also the documentation is not upfront in suggesting that it can predict. It says “The SDR classifier maps input patterns to class labels.” Since, I know that SDRClassifier is used in getting predictions out of TM it makes me think that the input should be predictiveCells instead of activeCells based on the documentation. Documentation does not say that it maps current input patterns to class labels that could occur in future.
I think that either the name of the class and/or the documentation should be explicit in clarifying that to the user.