Mechanism for refining values

[K-C-Martin]

Select the top (2^n)/(3^n) psets, where n is the number of parameters being varied...

Visualize selecting a single cube (defined by 2 points per side) within a larger cube (defined by 3 points - candidate values) per side (n is equivalent to 'dimensions').

Next round could choose 3 new values - in between and slightly outside of the 2 selected values.

How to fix a value? Identify when sorting psets by eRMSE has all psets with that value in a block at the top of the list?

[K-C-Martin]

Use a threshold of 'change in eRMSE' to identify top psets that show similar (good) fit?

[K-C-Martin]

Rather focus on relative improvement in eRMSE...? Choose a relative threshold based on all eRMSE values... improve threshold each round (or stop process when eRMSE stops improving?)... fix values when other values fall below the dynamic threshold...?

[K-C-Martin]

Keep eRMSE values calculated for previous psets that had 2 values make it through (and which had 3 new values generated between and flanking the 2 seed values)... include them in the next-round selection (allow 2 values to be chosen that come from successive rounds...).

e.g. keeping 1-2 and 1-3 allows it to be possible that the 3rd round could start with, say, 2-5 and 1-3 as seed values:

1-1
_________2-4

1-2-----1-2
__________________3-7
_________2-5-----2-5
__________________3-8-------3-8
1-3-----1-3------1-3-------1-3
__________________3-9
_________2-6

Would make sense to keep psets from previous rounds? All other parameter values would also carry through...

[K-C-Martin]

NB: Need to narrow each successive 3-value window? Including psets from previous rounds could do this automatically...

[K-C-Martin]

Rather than creating a window around two adjacent par values, rather keep them and place the third value in between them - narrow down faster.

To fix a value, maybe set a threshold difference between the outer values for each par (stop subsetting when this threshold is reached - leave all three fixed)

[K-C-Martin]

Subdividing windows defined by two adjacent par values allows some old psets to be reused? Best case still need to create (3^n - 2^n) new psets (e.g. where n = 2 parameters, narrowing down to a window defined by 2 adjacent values for each parameter will require an additional 5 psets to be defined).

2^n grows slowly compared to 3^n... may not be worth the added overhead to 'reuse' old psets rather than regenerating and running them through R3PG...