scale, prevent outlier, remove process noise for speedup #124
Replies: 2 comments 1 reply
-
Q2. I think people would readily accept sampling from the prior to generate a baseline. This is consistent with what people already do with multivariate sensitivity runs, even in data-free models. I think it needs to be easier though. People already underuse sensitivity analysis, because it adds an extra set of time-consuming tasks to the modeling process. So the difficult behavior change is to reallocate the distribution of time spent on the entire package. |
Beta Was this translation helpful? Give feedback.
-
Increasing N:10 -> 100 -> 200, fitting time increase 1-> 15s -> 35s (slight superlinear) |
Beta Was this translation helpful? Give feedback.
-
For the first topic (speedup) from #76 (comment) hier.ode.sbc blueprint, scaling, removing process noise structure, removing data series structure are tried
scaling
parameters are tight enough, but additive measurement noise is larger than other parameter as initial order 10000.
Q1. Will scaling this to 0~1 would make the fitting better in both accuracy and speed.
Long fitting time accompanies the following log:
Q2. will sd people ready to actually sample from prior distribution for baseline run? Most modeler's mental model seems to be using it as a penalty function for estimation
thin tail
either by
m_scale_sd
shape or measurement distribution. inven model time log:Q3. is it correct the second one is "likelihood to lognormal" formulation? what is the main difference between the two?
vs
process noise structure
is removed for 12W1 experiments on inventory model but the data series is still driving data (changed to sin wave)
Beta Was this translation helpful? Give feedback.
All reactions