Lightweight Adaptive Balancer (LUA version)

Given

a total evaluation budget N
an initial budget N0
a pruning factor of P,

this would explore (say) 10,000 unlabelled items as follows:

use N0 evalaution to label four items ;
divide labelled items into best and rest using a multi-goal criteria;
build a classifier to find, for unlabelled item, B=like(best), R=like(resti);
sort, in ascending order, the remaining 9996 items by -B/R;
label the top item (so N=N-1), and
discard the last P% unlabelled items;
update best and rest with this new item;
if N > 0, loop back to (2)
- Else terminate, returning the top item in best.