Currently, we first calculate some bitboards at the top of Evaluation::space()
and then check whether we actually need them. Invert the ordering. Of course this
does not make a difference in current master because the constexpr bitboard
calculations are in fact done at compile time by any decent compiler, but I find
my version a bit healthier since it will always meet or exceed current implementation
even if we eventually change the spaceMask to something not contsexpr.
No functional change.
Introduce voting system for best move selction in multi-threads mode.
Joint work with Stefan Geschwentner, based on ideas introduced by
Michael Stembera.
Moves are upvoted by every thread using the margin to the minimum score
across threads and the completed depth.
First thread voting for the winner move is selected as best thread.
Passed STC, LTC. A further LTC test with only 4 threads failed with positive
score. A LTC with 31 threads was stopped with LLR 0.77 after 25k games to
avoid use of excessive resources (equivalent to 1.5M STC games).
Similar ideas were proposed by Michael Stembera 2 years ago #507, #508.
This implementation seems simpler and more understandable, the results
slightly more promising.
Further possible work:
1) Tweak of the formula using for assigning votes.
2) Use a different baseline for the score dependent part: maximum score
or winning probability could make more sense.
3) Assign votes in `Thread::Search` as iterations are completed and use
voting results to stop search.
4) Select best thread as the threads voting for best move with the highest
completed depth or, alternatively, vote on PV moves.
Link to SPRT tests
[stopped LTC, 31 threads 20+0.02](http://tests.stockfishchess.org/tests/view/5b61dc090ebc5902bdb95192)
LLR: 0.77 (-2.94,2.94) [0.00,5.00]
Total: 25602 W: 3977 L: 3850 D: 17775
Elo: 1.70 [-0.68,4.07] (95%)
[passed LTC, 8 threads 20+0.02](http://tests.stockfishchess.org/tests/view/5b5df5180ebc5902bdb9162d)
LLR: 2.96 (-2.94,2.94) [0.00,5.00]
Total: 44478 W: 7602 L: 7300 D: 29576
Elo: 1.92 [-0.29,3.94] (95%)
[failed LTC, 4 threads 20+0.02](http://tests.stockfishchess.org/tests/view/5b5f39ef0ebc5902bdb92792)
LLR: -2.94 (-2.94,2.94) [0.00,5.00]
Total: 29922 W: 5286 L: 5285 D: 19351
Elo: 0.48 [-1.98,3.10] (95%)
[passed STC, 4 threads 5+0.05](http://tests.stockfishchess.org/tests/view/5b5dbf0f0ebc5902bdb9131c)
LLR: 2.97 (-2.94,2.94) [0.00,5.00]
Total: 9108 W: 2033 L: 1858 D: 5217
Elo: 6.11 [1.26,10.89] (95%)
No functional change (in simple threat mode)
[STC](http://tests.stockfishchess.org/tests/view/5b2614000ebc5902b8d17193)
LLR: 2.96 (-2.94,2.94) [-3.00,1.00]
Total: 17733 W: 3996 L: 3866 D: 9871
[LTC](http://tests.stockfishchess.org/tests/view/5b264d0f0ebc5902b8d17206)
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 55524 W: 9535 L: 9471 D: 36518
Use pawn count scaling also for opposite bishops endings with additional material, with a slope of 2 instead of 7. This simplifies slightly the code.
This PR is a functionally equivalent refactoring of the version which was submitted.
Four versions tried, 2 passed both STC and LTC. I picked the one which seemed more promising at LTC.
Slope 4 passed STC (-0.54 Elo), LTC not attempted
Slope 3 passed STC (+2.51 Elo), LTC (-0.44 Elo)
Slope 2 passed STC (+2.09 Elo), LTC (+0.04 Elo)
Slope 1 passed STC (+0.90 Elo), failed LTC (-3.40 Elo)
Bench: 4761613
This is a further step in the long quest for a simple way of determining
scale factors for the endgame.
Here we remove the artificial restriction in evaluate_scale_factor()
based on endgame score. Also SCALE_FACTOR_ONEPAWN can be simplified
away. The latter is a small non functional simplification with respect
to the version that was testedin the framework, verified on bench with
depth 22 for good measure.
Passed STC
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 49438 W: 9999 L: 9930 D: 29509
http://tests.stockfishchess.org/tests/view/5ae20c8b0ebc5963175205c8
Passed LTC
LLR: 2.96 (-2.94,2.94) [-3.00,1.00]
Total: 101445 W: 15113 L: 15110 D: 71222
http://tests.stockfishchess.org/tests/view/5ae2a0560ebc5902a1998986
How to continue from there?
Maybe the general case could be scaled with pawns from both colors
without losing Elo. If that is the case, then this could be merged
somehow with the scaling in evaluate_initiative(), which also uses
a additive malus down when the number of pawns in the position goes
down.
Closes https://github.com/official-stockfish/Stockfish/pull/1570
Bench: 5254862
Add a logarithmic term in the optimism computation, increase
the maximal optimism and lower the contempt offset.
This increases the dynamics of the optimism aspects, giving
a boost for balanced positions without skewing too much on
unbalanced positions (but this version will enter panic mode
faster than previous master when behind, trying to draw faster
when slightly behind). This helps, since optimism is in general
a good thing, for instance at LTC, but too high optimism
rapidly contaminates play.
passed STC:
LLR: 2.96 (-2.94,2.94) [0.00,5.00]
Total: 159343 W: 34489 L: 33588 D: 91266
http://tests.stockfishchess.org/tests/view/5a8db9340ebc590297cc85b6
passed LTC:
LLR: 2.97 (-2.94,2.94) [0.00,5.00]
Total: 47491 W: 7825 L: 7517 D: 32149
http://tests.stockfishchess.org/tests/view/5a9456a80ebc590297cc8a89
It must be mentioned that a version of the PR with contempt 0
did not pass STC [0,5]. The version in the patch, which uses
default contempt 12, was found to be as strong as current master
on different matches against SF7 and SF8, both at STC and LTC.
One drawback maybe is that it raises the draw rate in self-play
from 56% to 59%, giving a little bit less sensitivity for SF
developpers to find evaluation improvements by selfplay tests
in fishtest.
Possible further work:
• tune the values accurately, while keeping in mind the drawrate issue
• check whether it is possible to remove linear and offset term
• try to simplify the S-shape curve
Bench: 5934644
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.
STC:
LLR: 2.95 (-2.94,2.94) [0.00,5.00]
Total: 110052 W: 24614 L: 23938 D: 61500
http://tests.stockfishchess.org/tests/view/5a72e6020ebc590f2c86ea20
LTC:
LLR: 2.97 (-2.94,2.94) [0.00,5.00]
Total: 16470 W: 2896 L: 2705 D: 10869
http://tests.stockfishchess.org/tests/view/5a76c5b90ebc5902971a9830
A second match at LTC was organised against the current master:
ELO: 1.45 +-2.9 (95%) LOS: 84.0%
Total: 19369 W: 3350 L: 3269 D: 12750
http://tests.stockfishchess.org/tests/view/5a7acf980ebc5902971a9a2e
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.
Bench: 5207156
STC
LLR: 2.96 (-2.94,2.94) [-3.00,1.00]
Total: 58462 W: 10615 L: 10558 D: 37289
LTC
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 65061 W: 8539 L: 8477 D: 48045
It is worth noting that an attempt to only increase the bonus passed STC but failed LTC, and
an attempt to remove the cap without increasing the bonus is still running at STC, but will probably fail after more than 100k.
Bench: 6188591
Closes#1063
Minor non-functional simplifications in computing the scale factor.
In my opinion, the code is now slightly more readable:
- remove one condition which can never be satisfied.
- immediately return instead of assigning the sf variable.
Tested for non-regression:
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 62162 W: 11166 L: 11115 D: 39881
No functional change
Closes#992
After we have taken into account all cheap evaluation
terms, we check whether the score exceeds a given threshold.
If this is the case, we return a scaled down evaluation.
STC:
LLR: 3.35 (-2.94,2.94) [0.00,5.00]
Total: 12575 W: 2316 L: 2122 D: 8137
LTC:
LLR: 2.95 (-2.94,2.94) [0.00,5.00]
Total: 67480 W: 9016 L: 8677 D: 49787
Current version is the one rewritten by ceebo
further edited by me.
Bench: 5367704
In evaluate, we start by initializing the pos.psq_score
and adding the material imbalance. After that, we check
whether a specialized eval exists and if yes we return
that value and discard whatever we have computed until now.
It sounds more logical to first probe material entry and
return if we have a specialized eval, and only if it is
not the case initialize eval with some values. There is
no measurable speed-difference on my computer.
Non functional change.
Collect and give a second try to some almost passed tuning attempts and
one-line tweaks from the last month.
Passed STC
LLR: 3.07 (-2.94,2.94) [0.00,4.00]
Total: 15124 W: 2974 L: 2756 D: 9394
And LTC
LLR: 2.95 (-2.94,2.94) [0.00,4.00]
Total: 21577 W: 3507 L: 3289 D: 14781
Bench: 8855226
Resolves#464
