This seems a die hard idea :-)
Passed both short TC
LLR: 2.97 (-2.94,2.94) [-1.50,4.50]
Total: 17485 W: 3307 L: 3156 D: 11022
And long TC
LLR: 2.97 (-2.94,2.94) [0.00,6.00]
Total: 38181 W: 6002 L: 5729 D: 26450
bench: 8659830
Actually, it is not used, as both arrays have the
same values. Some local tests in either direction
showed no improvement.
Also some minor corrections in the comments.
No functional change.
Previously some squares could be "incorrectly" awarded
to a pinned piece.
e.g. in 3k4/1q6/3b4/3Q4/8/5K2/B7/8 b - - 0 1 the black
bishop get 4 squares too many and the white queen gets 6.
Passed both short TC.
LLR: 2.97 (-2.94,2.94) [-1.50,4.50]
Total: 4871 W: 934 L: 817 D: 3120
And long TC:
LLR: 2.96 (-2.94,2.94) [0.00,6.00]
Total: 38968 W: 6113 L: 5837 D: 27018
bench: 9282549
But compensate by reducing rook and queen
value by 53 = (160 / 3)
Material imbalances are affected as follows:
Red. Major Rook Queen Total
QRR +160 -2*53 -53 +1
QR +160 -53 -53 +54
RR +160 -2*53 0 +54
R 0 -53 0 -53
Q 0 0 -53 -53
so that the imbalance changes by at most 54 + 53 = 107 units.
This corresponds to appromximately 3.5cp in the final evaluation.
Verified with fixed number 40000 games at both short
and long TC it does not regress.
Short TC 15+0.05
ELO: 1.93 +-2.1 (95%) LOS: 96.6%
Total: 40000 W: 7520 L: 7298 D: 25182
Long TC 60+0.05
ELO: -0.33 +-1.9 (95%) LOS: 36.5%
Total: 39663 W: 6067 L: 6105 D: 27491
bench: 6703846
As, Gary (that analyzed the bug) says:
SF does not print a PV when the original best move fails low,
we hit our time allowance, and stop the search. The output from
the SF search is below. It was failing low on Ne1 at depth 34.
Then, we get bestmove Qd3, but no PV change.
info depth 34 seldepth 45 score cp 38 upperbound nodes 483484489 nps 15464575 time 31264 multipv 1 pv f3e1 h5h4 e1d3 h4g3 f2g3 a6f6 f1f6 e7f6 d1a4 f6e7 a1f1 d8f8 a4b3 b7b6 b3c2 f7f6 c2a4 h3g5 b2b3 g5f7 a4c6 f7d6 h1g2 f6f5 e4f5 d6f5
info depth 34 seldepth 45 score cp 38 upperbound nodes 483484489 nps 15464575 time 31264 multipv 1 pv f3e1 h5h4 e1d3 h4g3 f2g3 a6f6 f1f6 e7f6 d1a4 f6e7 a1f1 d8f8 a4b3 b7b6 b3c2 f7f6 c2a4 h3g5 b2b3 g5f7 a4c6 f7d6 h1g2 f6f5 e4f5 d6f5
info depth 34 seldepth 47 score cp 30 upperbound nodes 2112334132 nps 17255517 time 122415 multipv 1 pv f3e1 h5h4 d1a4 a6f6 e1d3 d8f8 a4c2 h4g3 f2g3 f6f1 a1f1 h7g8 b2b3 f7f6 a3a4 b7b6
info depth 34 seldepth 47 score cp 30 upperbound nodes 2112334132 nps 17255517 time 122415 multipv 1 pv f3e1 h5h4 d1a4 a6f6 e1d3 d8f8 a4c2 h4g3 f2g3 f6f1 a1f1 h7g8 b2b3 f7f6 a3a4 b7b6
info nodes 18235667001 time 969824
bestmove e2d3 ponder c8d7
Looking at the code, if we hit Signals.stop, we return from id_loop
before printing any PV. It is possible for us to have resorted the
RootMove list though, which will change the move that is actually
played.
No functional change.
1/ eval margin and gains removed:
16bit are now free on TT entries, due to the removal of eval margin. may be useful
in the future :) gains removed: use instead by Value(128). search() and qsearch()
are now consistent in this regard.
2/ futility_margin()
linear formula instead of complex (log(depth), movecount) formula.
3/ unify pre & post futility pruning
pre futility pruning used depth < 7 plies, while post futility pruning used
depth < 4 plies. Now it's always depth < 7.
Tested with fixed number of games both at short TC:
ELO: 0.82 +-2.1 (95%) LOS: 77.3%
Total: 40000 W: 7939 L: 7845 D: 24216
And long TC
ELO: 0.59 +-2.0 (95%) LOS: 71.9%
Total: 40000 W: 6876 L: 6808 D: 26316
bench 7243575
1/ eval margin and gains removed:
- gains removed by Value(128): search() and qsearch() now behave consistently!
2/ futility_margin()
- testing showed that there is no added value in this weird (log(depth), movecount)
formula, and a much simpler linear formula is just as good. In fact, it is most
likely better, as it is not yet optimally tuned.
- the new simplified formula also means we get rid of FutilityMargins[], its
initialization code, and more importantly ss->futilityMoveCount, and the hacky
code that updates it throughout the search().
- the current formula gives negative futility margins, and there is a hidden interaction
between the move coutn pruning formula and the futility margin one: what happens is
that MCP is supposed to be triggered before we use the non-sensical negative futility
margins.
3/ unify pre & post futility pruning
- pre futility pruning (what SF calls value based pruning) used depth < 7 plies,
while post futility pruning (what SF calls static null move pruning) used depth < 4 plies.
- also the condition depth < 7 in pre futility pruning was not obvious, and it seemd
to be depth < 16 (futility_margin() returns an infinite value when depth >= 7).
Tested with fixed number of games both at short TC:
ELO: 0.82 +-2.1 (95%) LOS: 77.3%
Total: 40000 W: 7939 L: 7845 D: 24216
And long TC
ELO: 0.59 +-2.0 (95%) LOS: 71.9%
Total: 40000 W: 6876 L: 6808 D: 26316
bench: 10206576
RedundantRook and RedundantQueen replaced by simple
variable RedundantMajor. Also the SameColor coefficient
for Queen<->Queen has been set by definition to 0.
The remaining 5 parameters:
LinearCoefficients[ROOK]
LinearCoefficients[QUEEN]
QuadraticCoefficientsSameColor[ROOK][ROOK]
QuadraticCoefficientsSameColor[QUEEN][ROOK]
RedundantMajor
are sufficient to equate the material imbalances for the
5 common material configurations of R, RR, Q, QR and QRR
to any desired values simultaneously.
With the chosen parameters there should be no functional
change unless one side has more than 2 rooks or more
than 1 queen. For example bench from the start position
using the commands:
./stockfish
go depth 16
produces identical output except for one extra node
in the last iteration.
bench: 8198094
Coefficients for Bishop<->BishopPair and Bishop<->Bishop
are also pretty much redundant. By altering the values
in LinearCoefficients[] these coefficients can be zeroed
without changing the imbalance calculations in any position
with less than 3 bishops for one side.
bench: 7995098
First coefficient in the SameColor array does an
equivalent job when folded into the LinearCoefficients
array.
All of the diagonal terms in the OppositeColor array
are redundant due to cancellation.
No functional change.
In case we find a very good move after a
troubled start, we don't return immediately
anymore.
Tested directly at long TC where it passed:
LLR: 2.95 (-2.94,2.94) [0.00,6.00]
Total: 13910 W: 2397 L: 2228 D: 9285
bench: 7995098
And remove a complex (and broken) formula.
Indeed previous code was broken in case of TC with big
time increments where available_time() was too similar
to total time yielding to many time losses, so for instance:
go wtime 2600 winc 2600
info nodes 4432770 time 2601 <-- time forfeit!
maximum search time = 2530 ms
available_time = 2300 ms
For a reference and further details see:
https://groups.google.com/forum/?fromgroups=#!topic/fishcooking/dCPAvQDcm2E
Speed tested with bench disabling timer alltogheter vs timer set at
max resolution, showed we have no speed regressions both in single
core and when using all physical cores.
No functional change.
A combo of two patches that failed SPRT with score
higher than 50% but togheter they succeed:
SPRT at 60+0.05
LLR: 2.95 (-2.94,2.94) [0.00,6.00]
Total: 7312 W: 1276 L: 1139 D: 4897
bench: 8029334
If the game got late enough that move_importance(currentPly) * slowMover / 100
rounds to 0, then we ended up dividing 0 by 0 when only looking 1 move ahead.
This apparently caused the search to almost immediately abort and Stockfish
would blunder in long games. So convert thisMoveImportance to a double.
No functional change.
This seems more a material imbalance topic,
anyhow test is good and so patch is applied
as is.
Passed both short TC:
LLR: 2.96 (-2.94,2.94) [-1.50,4.50]
Total: 17391 W: 3548 L: 3393 D: 10450
And long TC:
LLR: 3.00 (-2.94,2.94) [0.00,6.00]
Total: 34660 W: 5972 L: 5700 D: 22988
bench: 8291883
Dumb down a bit the code and trade some possible
speed (but this is far from hot path anyhow) for
some added readability for the layman.
No functional change.
The normalising transformation is computed all at
once by the helper function get_flip_sq and then
applied immediately to the relevant squares as soon
as they are loaded from the position class.
bench: 8350690
New formula mathces the old formula until d = 45
Test code:
int main() {
for(int d=1; d<=45; d++)
{
int a = int(log(double(d * d) / 2) / log(2.0) + 1.001);
int b = int(2.9 * log(double(d)));
if (a != b) std::cout << d << std::endl;
}
return 0;
}
bench: 8455956
This is the first chain bonus version
from Ralph that also passed both
Short TC:
LLR: 2.96 (-2.94,2.94) [-1.50,4.50]
Total: 23460 W: 4727 L: 4556 D: 14177
And long TC:
LLR: 2.95 (-2.94,2.94) [0.00,6.00]
Total: 31858 W: 5497 L: 5240 D: 21121
And performed better against current
committed version, always at 60secs:
LLR: -2.94 (-2.94,2.94) [-3.00,3.00]
Total: 26301 W: 4477 L: 4580 D: 17244
This test was done by Leonid.
bench: 8455956
After 40K games at 60 secs, result is still
not clear, but not a regression against SF 4
After
ELO: 50.11 +-2.1 (95%) LOS: 100.0%
Total: 40000 W: 10547 L: 4817 D: 24636
Before
ELO: 49.51 +-2.1 (95%) LOS: 100.0%
Total: 40000 W: 10483 L: 4821 D: 24696
So re-apply the patch to avoid to
special-case this one.
bench: 7403882
It was never updated !
Currently it only affects evaluate_passed_pawns()
and in particularly the rule to increase the bonus
if we have more non-pawn pieces. We could simply use
popcount() instead and avoid the little slowdown
in put_piece() and remove_piece(), but this would
leave a very subtle and tricky hole where people
are forced to remember that pos.count<ALL_PIECES>()
does not work. This is not obvious and so dangerous.
Thanks to Ronald de Man for spotting this.
bench: 7931424
Due to a strange issue (bug?) the ternary
operator does not return a BitCountType for
icc, so revert to the expression.
The same patch was already applied in
9749f1f14c
Thanks to NssY Wanyonyi for pointing out
this.
No functional change.
This reverts commit 4bc2374450 for
two reasons.
First regression testing shows almost equal
score:
Before the patch:
ELO: 49.75 +-2.5 (95%) LOS: 100.0%
Total: 27205 W: 7113 L: 3244 D: 16848
After the patch:
ELO: 48.87 +-2.9 (95%) LOS: 100.0%
Total: 20860 W: 5478 L: 2563 D: 12819
Second, and more sensible to me, this patch
increases safe check bonuses to 4 times their
original value (!) and considering:
- Values were already well tuned
- Values are highly critical
- King safety is highly critical, very TC
dependent and very difficult to test
- Our testing coverage is partial (self-testing,
blitz times)
I think is better to be safe than sorry and so
I revert the patch.
bench: 8440524
Passed both short TC:
LLR: 2.95 (-2.94,2.94) [-1.50,4.50]
Total: 10466 W: 2087 L: 1953 D: 6426
And long TC:
LLR: 2.96 (-2.94,2.94) [0.00,6.00]
Total: 26334 W: 4540 L: 4310 D: 17484
And also proved stronger than a slightly
different patch, also succesful against master:
https://github.com/mcostalba/Stockfish/commit/dc6830a3b4ed12
But losing against current one in a match
at 60secs with SPRT [-3, 3]:
LLR: -2.96 (-2.94,2.94) [-3.00,3.00]
Total: 44484 W: 7360 L: 7463 D: 29661
bench: 9160831
Also the drawing criteria has been slightly loosened.
It now detects a draw if the king is ahead of all the
pawns and on the same file or the adjacent file.
bench: 7700683
This lost position 8/8/3q4/8/5k2/2P1R3/2K2P2/8 w - - 0 1
was previously evaluated as a draw.
The king and rook need to be correctly placed with
respect to the _same_ pawn.
(Note also that the check for the pawn being on RANK_2
in the old version is redundant: it must be on RANK_2 if
it hopes to protect a rook on RANK_3)
Signed-off-by: Marco Costalba <mcostalba@gmail.com>
Good both at short TC:
LLR: 2.95 (-2.94,2.94) [-1.50,4.50]
Total: 5448 W: 1133 L: 1012 D: 3303
And at long TC:
LLR: 2.95 (-2.94,2.94) [0.00,6.00]
Total: 40509 W: 6836 L: 6541 D: 27132
bench: 7700683
Don't need a struct here. Speed test shows
result is teh same. Moreover RKISS is used
mainly at startup to compute magics, so
prefer to keep it simple...RKISS ;-)
Also some assorted triviality while there.
No functional change.
These two changes go in opposite directions and it
seems that the combination is stronger than original.
Here are the positive tests at various TC:
15+0.05
LLR: 2.96 (-2.94,2.94) [-1.50,4.50]
Total: 24561 W: 4946 L: 4772 D: 14843
60+0.05
LLR: 2.96 (-2.94,2.94) [0.00,6.00]
Total: 15259 W: 2598 L: 2423 D: 10238
40/30
LLR: 2.96 (-2.94,2.94) [-3.00,3.00]
Total: 2570 W: 527 L: 422 D: 1621
Unfortunately there is also a bad result
with one sec time increment that needs
to be further investigated:
12+1
LLR: -2.97 (-2.94,2.94) [-3.00,3.00]
Total: 2694 W: 438 L: 543 D: 1713
bench: 8340585
Rationale:
- Speed of double and float is about the same (not on the hot path anyway)
- Double makes code prettier (no need to write 1.0f, just 1.0)
- Only practical advantage of float is to use less memory, but since we never
store large arrays of double, we don't care.
No functional change.
