...
Users can set separate learning rates for four kinds of free parameters of variational posterior parameters: mean of node heights, standard deviation of node heights, mean of population sizes, and standard deviation of population sizes. These learning rates need to be set very carefully so that BBVI can converge quickly. During the BBVI search, VI_coalHMM will print the gradient of each parameter (mean and standard deviation of each demographic parameter) to the console. It is recommended that the user set the four learning rates according to the scale of the gradient of each parameter so that the step size of each parameter in each iteration is reasonable.
...