I thought of the following greedy, (but not optimal) algorithm to create a
pseudo-optimal prelude:
First we record the number of iterations (= states scanned) it took the scans'
to solve a given board for each of a large set of boards. (in our case the
Microsoft 32,000).
Then, we allocate a certain number of iterations, and assign this quota to
the scan that solves the most boards within this quota.
Repeat.
The configurations generated by this algorithm yield very good performance.