Sample more often to make better use of the computation time.
Beforehand, when starting from zero, convergence was far from
reached in the time the demo ran, while much time was used
to optimize for the initial samples.