Without reworking stuff there are two options on how to integrate parameter tuning - either via an update or via a measurement. I don't like either option. Both have an issue with how frequently they are applied. For updates the number of other updates may throttle the number of adaptive steps and for measurements the is the simulation-global measure_rate. So I might end up reworking a stuff again.
On the topic of the actual tuning I played around with a small attractive Hubbard model. In that case \kappa was pretty much always below 1 / sqrt(t). Since \kappa goes into the tuning inversely the tuning rate is always reduced. Reducing it at the start helps reducing jerky adaptions, but this is only necessary for a couple tens of sweeps. (I.e. not updating for the first 32 sweeps has a similar effect.) It might also make sense to throttle near the end of the simulation to smooth out the chemical potential values.
Closes #152
Without reworking stuff there are two options on how to integrate parameter tuning - either via an update or via a measurement. I don't like either option. Both have an issue with how frequently they are applied. For updates the number of other updates may throttle the number of adaptive steps and for measurements the is the simulation-global
measure_rate
. So I might end up reworking a stuff again.On the topic of the actual tuning I played around with a small attractive Hubbard model. In that case \kappa was pretty much always below
1 / sqrt(t)
. Since \kappa goes into the tuning inversely the tuning rate is always reduced. Reducing it at the start helps reducing jerky adaptions, but this is only necessary for a couple tens of sweeps. (I.e. not updating for the first 32 sweeps has a similar effect.) It might also make sense to throttle near the end of the simulation to smooth out the chemical potential values.