I dont think this is the case, and likely scaling the step_scale would not work. I am not completely sure you have the correct potential for Boltzmann machine written down, as I understand that the energy function should be a cross product between X and Y here? - which means you would have a 10000 x 10000 matrix and taking the sum. But if you are meaning to express that X and Y are random sample from 2 nodes, then you should take the expectation - which is why dividing by 10000 gives you the expected result (as pymc3 does the sum of the Potential term here).
1 Like