Single Comment

Well, after about 200 games the score has crept up to 57% for the Persians.
(Note that the download still uses the rules of your website!)
I am playing 40 moves/min, which makes the average game take 3 min. So I
can do 20 games/hour, 480 per day. But I am playing with ponder off, so it
takes only a singe core to play a game, and I have a dual. So I am running
two matches in parallel. :-)

On the other core I started the test with promotion to Colonel. It hardly
seems to matter. After 60 games, the Spartans are still leading by 70%. So
I guess we can scratch this idea, and I probably won't finish the run.

As the Persians are now winning, weakening the Warlord to a Crowned Bishop
was too much. When  look at the games, it is also awkward: the Spartans
have very much trouble in the standard setup to develop it. Fairy-Max is
reluctant to move away the f-Pawn (Hoplite) because it is shielding a King
and also reluctant to play Hg7h6, because it puts a Hoplite on the edge.
And after Hg7f6 it still blocks the Warlord after Wg7. So apart from the
fact that the intrinsic difference between B+N and B+K is much larger than
that between R+N and R+K, it also does not make good use of the B+K.

So the next thing I will try (on the other core) is to keep Warlord as B+N,
and make the General R+K. This should work much more smoothly: moving the
b-Hoplite immediately opens a file for it, so there should be no problem
developing. Promotion can then be to B+N Warlord, which, based on the test
with Colonel-promotion, should cause no weakening at all. An alternative would be to equip the B+K piece with some extra moves that help in development, e.g. a (2,2)-jump, or when that is too strong, only as a non-capture. But that makes it a bit of a weird piece, so I would prefer the clean R+K if it does the job.

When I have a setup that is fairly balanced, I will start running tests
with some pieces deleted from the opening array, (preferably in pairs, to
increase sensitivity), like Colonels on one side and Knights on the other,
to see how this changes the score (i.e. to whom such a trade would be
favorable). This to verify if the values I programmed into Fairy-Max are
close enough to reality. If not, it would be wise to repeat the tests with
corrected piece values in a second iteration. My experience is that
slightly wrong values hardly affect the score, as long as both sides share
the same misconception. (E.g. it does not matter if you program the value
of the Bishop higher or lower than that of the Knight: two Bishops will
beat two Knights by ~58% in both cases.)