Im thinking about applying HTM to Q-learning similar to how Neural networks is applied to Q-learning(Deep Q-networks). But I have some questions before exploring this further.
Can HTM do scalar predictions?
And if so, how does it perform regression relative to Neural networks on non-temporal data & non image data?
Is the HTM output limited to a range(e.g. [0, 100])? Does it perform worse for very large scalar predictions > 1000.
Can it do multi-scalar predictions?
Yes.
It does not perform regression. The mechanism is entirely different. Read BAMI or watch HTM School.
If the output is a prediction, it is generally going to be within the range of the input data.
Do you mean 1000 steps ahead in time? If then, yes. If you just mean large scalar values, it all depends on how you encode your input. If your typical min/max is 1/100 and you see 1000, it is typically clipped, but it depends on your encoder settings. See this tutorial.
It is possible, but needs work: Predicting Multiple Output Values