factor_sequence
For under-relaxation of shared sample increments, apply a sequence of factors, one per iteration
Specification
Alias: None
Arguments: REALLIST
Description
When initial correlation / covariance approximations are inaccurate, iterated sample updates can be overly aggressive, such that it becomes advisable to only accept a portion of this suggested increment through use of an under-relaxation factor.
In the factor_sequence
option, a sequence of relaxation factors is
specified, one for each online iteration. Once the sequence is
exhausted, all subsequent online iterations will continue to employ
the final sequence value.
Usage Tips
While not required or enforced, it is often a good idea to end the sequence with a value of 1, corresponding to retirement of the relaxation process.
The one case where the relaxation factor is automatically advanced to
1 is the case where a max_iterations
constraint will force early
termination of the iteration. This avoids polluting the final results
with an anticipated variance reduction that was not fully realized due
to increment throttling.
The need for under-relaxation is usually highest on the initial iteration, corresponding to a smaller relaxation factor in the sequence (e.g., initial factor of .5 below). This need then decays as the shared samples approach the optimal allocation and the covariances approach their final values (e.g., subsequent factors of .75 and 1. below).
Examples
method,
multilevel_blue
solution_mode online_pilot
relaxation factor_sequence = .5 .75 1.
pilot_samples = 25
seed = 8674132
max_function_evaluations = 500