ExpRegularization
phasic.svgd.ExpRegularization(first_reg=1.0, last_reg=0.0, tau=1000.0)Exponential decay schedule: reg = first_reg * exp(-iteration/tau) + last_reg * (1 - exp(-iteration/tau)).
This schedule helps by starting with strong moment regularization to guide initial exploration, then gradually reducing regularization as optimization converges to allow fine-tuning.
Parameters
first_reg :float= 1.0-
Initial (first) regularization strength at iteration 0
last_reg :float= 0.0-
Final (last) regularization strength as iteration → ∞
tau :float= 1000.0-
Decay time constant (larger = slower decay)
Examples
>>> schedule = ExpRegularization(first_reg=5.0, last_reg=0.1, tau=500.0)
>>> schedule(0) # iteration 0
5.0
>>> schedule(500) # iteration 500 (≈63% decay)
0.1925
>>> schedule(5000) # iteration 5000 (full decay)
0.1