It’s not obvious from this explanation how the tune parameter is calibrated. Could someone explain?
If the acceptance rate is below target (1/3 in the default), the algorithm will decrease the scaling factor that multiplies the proposal covariance, if it’s above target, the scaling factor will be increases. The update takes place every 500 draws.
Interesting, and by how much does the scale parameter jump if above or below the target? Does the algorithm target only the acceptance rate over the past 500 draws? And does it do this through both MCMCs or just the initial one?
You can see the approach starting at line 124 of