Make contempt dependent on the current score of the root position.
The idea is that we now use a linear formula like the following to decide
on the contempt to use during a search :
contempt = x + y * eval
where x is the base contempt set by the user in the "Contempt" UCI option,
and y * eval is the dynamic part which adapts itself to the estimation of
the evaluation of the root position returned by the search. In this patch,
we use x = 18 centipawns by default, and the y * eval correction can go
from -20 centipawns if the root eval is less than -2.0 pawns, up to +20
centipawns when the root eval is more than 2.0 pawns.
To summarize, the new contempt goes from -0.02 to 0.38 pawns, depending if
Stockfish is losing or winning, with an average value of 0.18 pawns by default.
Finally, we checked that there is no apparent problem with multithreading,
despite the fact that some threads might have a slightly different contempt
level that the main thread.
Match of this version against master, both using 5 threads, time control 30+0.3:
ELO: 2.18 +-3.2 (95%) LOS: 90.8%
Total: 14840 W: 2502 L: 2409 D: 9929
http://tests.stockfishchess.org/tests/view/5a7bf3e80ebc5902971a9aa2
Include suggestions from Marco Costalba, Aram Tumanian, Ronald de Man, etc.
Leonid Pechenik [Tue, 6 Feb 2018 10:14:55 +0000 (05:14 -0500)]
Retire "Extra thinking before accepting draw PVs"
This patch simplifies the time management code, removing the extra
thinking time for moves with draw PV and increasing thinking time
for all moves proportionally by around 4%.
Last time when the time management was carefully tuned was 1.5-2 years
ago. As new patches were getting added, time management was drifting out
of optimum. This happens because when search becomes more precise pv and
score are becoming more stable, there are less fail lows, best move is
picked earlier and there are less best move changes. All this factors are
entering in time management, and average time per move is decreasing with
more and more good patches. For individual patches such effect is small
(except some) and may be up or down, but when there are many of them,
effect is more substantial. The same way benchmark with more and more
patches is slowly drifting down on average.
So my understanding that back in October adding more think time for draw
PV showed positive Elo because time management was not well tuned, there
was more time available, and think_hard patch applied this additional time
to moves with draw PV, while just retuning back to optimum would recover Elo
anyway. It is possible that absence of contempt also helped, as SF9 is showing
less 0.0 scores than the October version.
Anyway, to me it seems that proper place to deal with draw PV is search, and
contempt sounds as much better solution. In time management there is little
additional elo, and if some code is not helping like removed here, it is better
to discard it. It is simpler to find genuine improvement if code is clean.
• Passed an additional non-regression [-5..0] test at the time control
of 60sec for the game (sudden death) with disabled draw adjudication:
LLR: 2.95 (-2.94,2.94) [-5.00,0.00]
Total: 8438 W: 1675 L: 1586 D: 5177
http://tests.stockfishchess.org/tests/view/5a7c3d8d0ebc5902971a9ac0
• Passed an additional non-regression [-5..0] test at the time control
of 1sec+1sec per move with disabled draw adjudication:
LLR: 2.97 (-2.94,2.94) [-5.00,0.00]
Total: 27664 W: 5575 L: 5574 D: 16515
http://tests.stockfishchess.org/tests/view/5a7c3e820ebc5902971a9ac3
This is a functional change for the time management code.
The 'eval' debugging command in Terminal did not initialize the Eval::Contempt
variable, leading to random output during debugging sessions (normal search
was unaffected by the bug).
Example of session where the two 'eval' commands should give the same output,
but did not:
./stockfish
position startpos
d
eval
go depth 20
d
eval
The bug is fixed by initializing Eval::Contempt to SCORE_ZERO in Eval::trace
protonspring [Thu, 8 Feb 2018 09:45:45 +0000 (10:45 +0100)]
Obey skipQuiets strictly in MovePicker
The current logic in master is to continue return quiet moves if their
history score is above 0. It appears as though this check can be
removed, which is also more logically consistent with the “skipQuiets”
semantics used in search.cpp.
This patch may open new opportunitiesto get Elo by changing or
tuning the definition of 'moveCountPruning' in line 830 of search.cpp,
because obeying skipQuiets without checking the history scores makes
the search more sensitive to 'moveCountPruning'.
protonspring [Mon, 5 Feb 2018 16:27:59 +0000 (17:27 +0100)]
Don’t score and sort all captures in RECAPTURES stage.
For these recaptures, we’re are only considering those captures
that recapture the recapture square (small portion of all the
captures). Therefore, scoring all of the captures and pick_besting
out of the whole group is not necessary.
protonspring [Sun, 4 Feb 2018 13:57:57 +0000 (14:57 +0100)]
Simplify qsearch stages in MovePicker
The difference between QCAPTURES_1 and QCAPTURES_2 quiescence search stages
boils down to a simple check of depth. The way it's being done now is
unnecessarily complex.
This patch is simpler, clearer, and easier to understand.
Passed SPRT[-3..1] test at STC:
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 99755 W: 22158 L: 22192 D: 55405
http://tests.stockfishchess.org/tests/view/5a71f41c0ebc590f2c86e9cb
joergoster [Wed, 31 Jan 2018 17:23:57 +0000 (18:23 +0100)]
Implement 'old' multipv search.
It seems to be a waste of time to loop through all remaining root moves
after finishing each PV line. This patch skips this until we have reached
the last PV line (this is the way it was done in Glaurung and very early
versions of Stockfish).
syzygy1 [Sun, 28 Jan 2018 13:40:07 +0000 (14:40 +0100)]
This modifies the in-search TB probing to continue searching for a mate "behind" a TB win (or loss). (#1285)
It does the following:
- If a TB win or loss value allows an alpha or beta cutoff, the cutoff is taken.
- Otherwise, the search of the current subtree continues. In PV nodes, the final value returned is adjusted to reflect that the position is a TB win (or loss).
The patch also fixes a potential problem caused by root_probe() and root_probe_wdl() dirtying the root-move scores.
This patch removes the limitation of current master that a mate is never found if the root position is not yet in the TBs, but the path to mate at some point enters the TBs. The patch is intended to preserve the efficiency and effectiveness of the current TB probing approach.
Set the default contempt value of Stockfish to 20 centipawns.
The contempt feature of Stockfish tries to prevent the engine from
simplifying the position too quickly when it feels that it is very
slightly behind, instead keeping the tension a little bit longer.
Various tests in November 2017 have proved that our current imple-
mentation works well against SF7 (which is about 130 Elo weaker than
current master) and than the Elo gain is an increasing function of
contempt, going (against SF7) from +0 Elo when contempt is set at
zero centipawns, to +30 Elo when contempt is 40 centipawns.
This november discussion left open the decision of which "default"
value for contempt we should use for Stockfish, taking into account
the various uses ofStockfish (opening preparation for humans, computer
online tournaments,analysis tool for web pages, human/computer play,
etc).
This pull request proposes to set the default contempt value of SF
to twenty centipawns, which turns out to be the highest value which
is not a regression against current master, as this seemed to be a
good compromise between risk and safety. A couple of SPRT[-3..1]
tests were done to bisect this value:
Surprisingly, a test at "very long time control" hinted that using
contempt 20 is not only be non-regressive against contempt 0, but
may actually exhibit some small Elo gain, giving a likehood of superio-
rity of 88.7% after 8500 games:
See the comments in pull request 1361 for the long, nice discussion
(180 entries :-)) leading to the decision to propose contempt 20 as
the default value:
Whether Stockfish should strictly adhere to the Komodo and Houdini
semantics and add the UCI commands to force the contempt to be White
in the so-called "analysis mode" is still under discussion, and may
be or may not be the object of a future commit.
Günther Demetz [Sat, 13 Jan 2018 08:01:23 +0000 (09:01 +0100)]
Simplify verification search (#1362)
1. avoid recursive call of verification.
For the interested side to move recursion makes no sense.
For the other side it could make sense in case of mutual zugzwang,
but I was not able to figure out any concrete problematic position.
Allows the removal of 2 local variables.
2. avoid further reduction by removing R += ONE_PLY;
Benchmark with zugzwang-suite (see #1338), max 45 secs per position:
Patch solves 33 out of 37
Master solves 31 out of 37
as discussed in issue #1349, the way pages are allocated with calloc might imply some overhead on first write.
This overhead can be large and slow down the first search after a TT resize significantly, especially for large TT.
Using an explicit clear of the TT on resize fixes this problem.
Not implemented, but possibly useful for large TT, is to do this zero-ing using all search threads. Not only would this be faster, it could also lead to a more favorable memory allocation on numa systems with a first touch policy.
Upon changing the number of threads, make sure all threads are bound
The heuristic to avoid thread binding if less than 8 threads are requested resulted in the first 7 threads not being bound.
The branch was verified to yield a roughly 13% speedup by @CoffeeOne on the appropriate hardware and OS, and an earlier version of this patch tested well on his machine:
To make sure all threads (including mainThread) are bound as soon as the total number exceeds 7, recreate all threads on a change of thread number.
To do this, unify Threads::init, Threads::exit and Threads::set are unified in a single Threads::set function that goes through the needed steps.
The code includes several suggestions from @joergoster.
Allow for general transposition table sizes. (#1341)
For efficiency reasons current master only allows for transposition table sizes that are N = 2^k in size, the index computation can be done efficiently as (hash % N) can be written instead as (hash & 2^k - 1). On a typical computer (with 4, 8... etc Gb of RAM), this implies roughly half the RAM is left unused in analysis.
This issue was mentioned on fishcooking by Mindbreaker:
http://tests.stockfishchess.org/tests/view/5a3587de0ebc590ccbb8be04
Recently a neat trick was proposed to map a hash into the range [0,N[ more efficiently than (hash % N) for general N, nearly as efficiently as (hash % 2^k):
namely computing (hash * N / 2^32) for 32 bit hashes. This patch implements this trick and now allows for general hash sizes. Note that for N = 2^k this just amounts to using a different subset of bits from the hash. Master will use the lower k bits, this trick will use the upper k bits (of the 32 bit hash).
There is no slowdown as measured with [-3, 1] test:
1) the patch is implemented for a 32 bit hash (so that a 64 bit multiply can be used), this effectively limits the number of clusters that can be used to 2^32 or to 128Gb of transpostion table. That's a change in the maximum allowed TT size, which could bother those using 256Gb or more regularly.
2) Already in master, an excluded move is hashed into the position key in rather simple way, essentially only affecting the lower 16 bits of the key. This is OK in master, since bits 0-15 end up in the index, but not in the new scheme, which picks the higher bits. This is 'fixed' by shifting the excluded move a few bits up. Eventually a better hashing scheme seems wise.
Despite these two caveats, I think this is a nice improvement in usability.
Current master can yield different staticEvals depending on the path
used to reach the position. The reason for this is that the evaluation after a
null move is always computed subtracting 2 * Eval::Tempo, while this is not
the case for lazy or specialized evals. This patch always adds tempo to evals,
which doesn't affect playing strength:
Rocky640 [Sun, 17 Dec 2017 07:50:45 +0000 (02:50 -0500)]
Simplify other checks (#1337)
Replace an intricate definition with a more natural one.
Master was excluding squares occupied by a pawn which was blocked by a pawn.
This version excludes any squares occupied by a pawn which is blocked by "something"
Alain SAVARD [Sun, 10 Dec 2017 15:13:28 +0000 (10:13 -0500)]
Simplify other checks #1334
Simplify the other check penalty computation. Compared to current master,
a) it uses a 143 kingDanger penalty instead of S(10, 10) for the "otherCheck"
(credits to ElbertoOne for finding a suitable kingDanger range to replace the score
and to Guardian for showing this could also be a neutral change at LTC).
This makes our king safety model more consistent and simpler.
b) it might also score more than one "otherCheck" penalty for a given piece type instead of just one
c) it might score many pinned penalties instead of just one.
d) It also remove 3 conditionals and uses simpler expressions.
So it was tested as a SPRT[-3, 1]
Trying to improve on b) another attempt was made to score also the
"otherchecks" for piece types which had some safe checks, but this
failed STC http://tests.stockfishchess.org/tests/view/5a2c79e60ebc590ccbb8badd
A better contempt implementation for Stockfish (#1325)
* A better contempt implementation for Stockfish
The round 2 of TCEC season 10 demonstrated the benefit of having a nice contempt implementation: it gives the strongest programs in the tournament the ability to slow down the game when they feel the position is slightly worse, prefering to stay in a complicated (even if slightly risky) middle game rather than simplifying by force into a drawn endgame.
The current contempt implementation of Stockfish is inadequate, and this patch is an attempt to provide a better one.
Passed STC non-regression test against master:
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 83360 W: 15089 L: 15075 D: 53196
http://tests.stockfishchess.org/tests/view/5a1bf2de0ebc590ccbb8b370
This contempt implementation is showing promising results in certains situations. For instance, it obtained a nice +30 Elo gain when playing with contempt=40 against Stockfish 7, compared to current master:
• master against SF 7 (20000 games at LTC): +121.2 Elo
• this patch with contempt=40 (20000 games at LTC): +154.11 Elo
This was the result of real cooperative work from the Stockfish team, with key ideas coming from Stefan Geschwentner (locutus2) and Chris Cain (ceebo) while most of the community helped with feedback and computer time.
In this commit the bench is unchanged by default, but you can test at home with the new contempt in the UCI options. The style of play will change a lot when using contempt different of zero (I repeat: not done in this version by default, however)!
The Stockfish team is still deliberating over the best default contempt value in self-play and the best contempt modeling strategy, to help users choosing a contempt value when playing against much weaker programs. These informations will be given in future commits when available :-)
syzygy1 [Mon, 4 Dec 2017 16:52:31 +0000 (17:52 +0100)]
Use a Direction enum for Square deltas
Currently the NORTH/WEST/SOUTH/EAST values are of type Square, but conceptually they are not squares but directions. This patch separates these values into a Direction enum and overloads addition and subtraction to allow adding a Square to a Direction (to get a new Square).
I have also slightly trimmed the possible overloadings to improve type safety. For example, it would normally not make sense to add a Color to a Color or a Piece to a Piece, or to multiply or divide them by an integer. It would also normally not make sense to add a Square to a Square.
Add the -fno-exceptions flag to the Makefile to avoid the unecessary exceptions support in the executable (we do not use any exception in Stockfish at the moment).
This change gives a 9.2% reduction in size for the executable binary.
syzygy [Sat, 25 Nov 2017 22:20:36 +0000 (23:20 +0100)]
Minor cleanup of search.cpp
Four very minor edits. Note that tte->save() uses posKey and
not pos.key() in other places.
Originally I also added a futility_move_counts() function to
make things more consistent with the futility_margin() and
reduction() functions. But then razor_margin[] should probably
also be turned into a function, etc. Maybe a good idea, maybe not.
So I did not include it.
Rocky640 [Sat, 11 Nov 2017 12:37:29 +0000 (07:37 -0500)]
Simplify some kingring penalties expressions
The new "weak" expression helps simplify the safe check calculations for rooks or minors, (but the end result for all the safe checks is the exactly the same as in current master)
The only functional change is for the "outer king ring" (for example, squares f3 g3 h3 when white king is on g1). In current master, there was a 191 penalty if any of these was not defended at all.
With this pr, there is this 191 penalty if any of these is not defended at all or is only defended by a white queen.
Tested as a simplification
STC
http://tests.stockfishchess.org/tests/view/59fb03d80ebc590ccbb89fee
LLR: 2.96 (-2.94,2.94) [-3.00,1.00]
Total: 66167 W: 12015 L: 11971 D: 42181
(against master (Update Copyright year inMakefile))
LTC
http://tests.stockfishchess.org/tests/view/5a0106ae0ebc590ccbb8a55f
LLR: 2.96 (-2.94,2.94) [-3.00,1.00]
Total: 15790 W: 2095 L: 1968 D: 11727
(against master (Handle BxN trade as good capture when history scor))
same as #1296 but rebased on latest master
bench: 5109559
ceebo [Wed, 8 Nov 2017 17:21:46 +0000 (17:21 +0000)]
Add comments to pos.see_ge()
In terms of technical changes this patch eliminates the return
statements from the main loop of pos.see_ge() and replaces two conditional
computations with a single bitwise negation.
Stockfish currently relies on the "filter_root_moves" function also
having the side effect of clamping Cardinality against MaxCardinality
(the actual piece count in the tablebases). So if we skip this function,
we will end up probing in the search even without tablebases installed.
We cannot bail out of this function before this check is done, so move
the MultiPV hack a few lines below.
IIvec [Sun, 15 Oct 2017 14:44:29 +0000 (16:44 +0200)]
Fix premature using of all available time in x/y TC
In x/y time controls there was a theoretical possibility
to use all available time few moves before the clock will
be updated with new time. This patch fixes that issue.
The first change (ss->statScore >= 0) does nothing.
The second change ((ss-1)->statScore >= 0 ) has a massive change.
(ss-1)->statScore is not set until (ss-1) begins to apply LMR to moves.
So we now increase the reduction for bad quiets when our opponent is
running through the first captures and the hash move.
Let ss->ply denote the number of plies from the root to the current node
This patch lets ss->ply be equal to 0 at the root of the search.
Currently, the root has ss->ply == 1, which is less intuitive:
- Setting the rootNode bool has to check (ss-1)->ply == 0.
- All mate values are off by one: the code seems to assume that mated-in-0
is -VALUE_MATE, mate-1-in-ply is VALUE_MATE-1, mated-in-2-ply is VALUE_MATE+2, etc.
But the mate_in() and mated_in() functions are called with ss->ply, which is 1 in
at the root.
- The is_draw() function currently needs to explain why it has "ply - 1 > i" instead
of simply "ply > i".
- The ss->ply >= MAX_PLY tests in search() and qsearch() already assume that
ss->ply == 0 at the root. If we start at ss->ply == 1, it would make more sense to
go up to and including ss->ply == MAX_PLY, so stop at ss->ply > MAX_PLY. See also
the asserts testing for 0 <= ss->ply && ss->ply < MAX_PLY.
The reason for ss->ply == 1 at the root is the line "ss->ply = (ss-1)->ply + 1" at
the start for search() and qsearch(). By replacing this with "(ss+1)->ply = ss->ply + 1"
we keep ss->ply == 0 at the root. Note that search() already clears killers in (ss+2),
so there is no danger in accessing ss+1.
I have NOT changed pv[MAX_PLY + 1] to pv[MAX_PLY + 2] in search() and qsearch().
It seems to me that MAX_PLY + 1 is exactly right:
- MAX_PLY entries for ss->ply running from 0 to MAX_PLY-1, and 1 entry for the
final MOVE_NONE.
I have verified that mate scores are reported correctly. (They were already reported
correctly due to the extra ply being rounded down when converting to moves.)
The value of seldepth output to the user should probably not change, so I add 1 to it.
(Humans count from 1, computers from 0.)
A small optimisation I did not include: instead of setting ss->ply in every invocation
of search() and qsearch(), it could be set once for all plies at the start of
Thread::search(). This saves a couple of instructions per node.
No functional change (unless the search searches a branch MAX_PLY deep), so bench
does not change.
Do not use the opposed flag for scoring backward and isolated pawns
in pawns.cpp, instead give a S(5,25) bonus for each opponent unopposed
weak pawns when we have a rook or a queen on the board.
STC run stopped after 113188 games:
LLR: 1.63 (-2.94,2.94) [0.00,5.00]
Total: 113188 W: 20804 L: 20251 D: 72133
http://tests.stockfishchess.org/tests/view/59b58e4d0ebc5916ff64b12e
In light of issue #1232, a test was performed about the value of '-fno-exceptions' and a second one of the combination '-fno-exceptions -fno-rtti'. It turns out these options are can be removed without introducing slowdown.
Prevent Stockfish from exiting if DTZ table is not present
During TB initialisation, Stockfish checks for the presence of WDL
tables but not for the presence of DTZ tables. When attempting to probe
a DTZ table, it is therefore possible that the table is not present.
In that case, Stockfish should neither exit nor report an error.
To verify the bug:
$ ./stockfish
setoption name SyzygyTable value <path_to_WDL_dir>
position fen 8/8/4r3/4k3/8/1K2P3/3P4/6R1 w - -
go infinite
Could not mmap() /opt/tb/regular/KRPPvKR.rtbz
$
(On my system, the WDL tables are in one directory and the DTZ tables
in another. If they are in the same directory, it will be difficult
to trigger the bug.)
The fix is trivial: check the file descriptor/handle after opening
the file.
If any thread found a 'mate in x' stop the search. Previously only
mainThread would do so. Requires the bestThread selection to be
adjusted to always prefer mate scores, even if the search depth is less.
I've tried to collect some data for this patch. On 30 cores, mate finding
seems 5-30% faster on average. It is not so easy to get numbers for this,
as the time to find a mate fluctuates significantly with multi-threaded runs,
so it is an average over 100 searches for the same position. Furthermore,
hash size and position make a difference as well.
Use less reduction for moves with larger moveCount if your
opponent did an unexpected (== high moveCount) move in the
previous ply... unexpected moves might need unexpected answers.
lucasart [Sun, 20 Aug 2017 11:59:46 +0000 (19:59 +0800)]
Restore safety margin of 60ms
What this patch does is:
* increase safety margin from 40ms to 60ms. It's worth noting that the previous
code not only used 60ms incompressible safety margin, but also an additional
buffer of 30ms for each "move to go".
* remove a whart, integrating the extra 10ms in Move Overhead value instead.
Additionally, this ensures that optimumtime doesn't become bigger than maximum
time after maximum time has been artificially discounted by 10ms. So it keeps
the code more logical.
Marco Costalba [Tue, 15 Aug 2017 08:05:22 +0000 (01:05 -0700)]
Restore perft
Rewrite perft to be placed naturally inside new
bench code. In particular we don't have special
custom code to run perft anymore but perft is
just a new parameter of 'go' command.
Marco Costalba [Wed, 16 Aug 2017 09:28:54 +0000 (02:28 -0700)]
Speed up Trevis CI
Avoid a couple of redundant rebuilds and compile
with 2 threads since travis gives 2vCPUs.
Also enable -O1 optimization for valgrind and
sanitizers, it should be safe withouth false
positives and it gives a very sensible speed
up, especially with valgrind.
The spee dup allow us to increase testing to
depth 10, useful for thread sanitizer.
projects / stockfish / log