[modeller_usage] Lucky16: Ordered search across multiple models?

[R K Belew <>]

hi Ben,  i'm finding an odd pattern across what i think of as random
trials, and i wonder why.

Suppose i run N independently seeded runs from the same alignment and 
generate K models each time; ie,  i have a distribution of N * K models. 
 Now consider a small fraction EPSILON of these with good GA341
scores.  i would expect the fraction of models which happened to
be model "target.B ..._k"  ( ie, the model that gets generated the
k-th time by Modeller) to be uniform over choice of k, wouldn't you?

instead, i find that for some values of k, ~50% have very good GA341
scores (across various random seed instances), while some have none.
here's a sample distribution with K=20:

> ModNum	Freq
> 1	0.01
> 3	0.03
> 4	0.02
> 5	0.01
> 6	0.09
> 7	0.01
> 8	0.01
> 9	0.05
> 11	0.06
> 12	0.08
> 15	0.01
> 16	0.50
> 19	0.05
> 20	0.01

what makes model#16 so consistently good?1

	rik