Moderators: hgm, Rebel, chrisw
It is incredibly hard to improve an engine at this level. But in line with your comments, CCRL Blitz single CPU (2'+1" games) does show steady improvement
Code: Select all
Stockfish 15.1 64-bit 3724 +14 −14
Stockfish 15 64-bit 3703 +11 −11
Stockfish 14.1 64-bit 3692 +13 −13
Stockfish 14 64-bit 3665 +15 −15
Stockfish 13 64-bit 3639 +14 −14
Stockfish is optimized and strong enough, if he can play always the best move with the correct evaluation and the correct wdl evaluation within 1 second.Modern Times wrote: ↑Sun Dec 18, 2022 3:42 pmIt is incredibly hard to improve an engine at this level. But in line with your comments, CCRL Blitz single CPU (2'+1" games) does show steady improvement
Code: Select all
Stockfish 15.1 64-bit 3724 +14 −14 Stockfish 15 64-bit 3703 +11 −11 Stockfish 14.1 64-bit 3692 +13 −13 Stockfish 14 64-bit 3665 +15 −15 Stockfish 13 64-bit 3639 +14 −14
Code: Select all
1 stockfish15.1 +net +4 +148/=716/-136 50.60% 506.0/1000
2 stockfish15.1 -4 +136/=716/-148 49.40% 494.0/1000
?Why?Jouni wrote: ↑Tue Dec 20, 2022 5:52 pm Bad release timing? There is now new net nn-335a9b2d8a80.nnue, which gave +3 Elo in framework! My short test gave also gainSoon SF 15.1.1Code: Select all
1 stockfish15.1 +net +4 +148/=716/-136 50.60% 506.0/1000 2 stockfish15.1 -4 +136/=716/-148 49.40% 494.0/1000
If anything, a new net after 15.1 is good news so Stockfish gets a little bit closer to 15.2. Also, 15.1.1 would be a bugfix release.Jouni wrote: ↑Tue Dec 20, 2022 5:52 pm Bad release timing? There is now new net nn-335a9b2d8a80.nnue, which gave +3 Elo in framework! My short test gave also gainSoon SF 15.1.1Code: Select all
1 stockfish15.1 +net +4 +148/=716/-136 50.60% 506.0/1000 2 stockfish15.1 -4 +136/=716/-148 49.40% 494.0/1000
Thank god we have you throwing your magic dice that show you how to tweak the Stockfish parameters for a real real strong VVVVLTC engine.mehmet123 wrote: ↑Wed Dec 21, 2022 10:02 pm After Stockfish 15, ~ 30 green patches had been added to Stockfish. Elo progress is ~6 elo at VLTC( 8 cores /30 sec +0.3) tests. Elo progress at rapid/ long time games most likely below 5 elo. If all these green patches are retested within at least 5 minutes of test condition, patches that weaken Stockfish may be detected as game time increases.
Code: Select all
Nur in SRC-1.1/src: benchmark.o.
Nur in SRC-1.1/src: bitbase.o.
Nur in SRC-1.1/src: bitboard.o.
Nur in SRC-1.1/src: endgame.o.
diff -ur SRC-1.0a/src/evaluate.h SRC-1.1/src/evaluate.h
--- SRC-1.0a/src/evaluate.h 2022-11-19 21:38:06.000000000 +0100
+++ SRC-1.1/src/evaluate.h 2022-12-21 17:08:53.000000000 +0100
@@ -39,7 +39,7 @@
// The default net name MUST follow the format nn-[SHA256 first 12 digits].nnue
// for the build process (profile-build and fishtest) to work. Do not change the
// name of the macro, as it is used in the Makefile.
- #define EvalFileDefaultName "nn-154550620475.nnue"
+ #define EvalFileDefaultName "nn-335a9b2d8a80.nnue"
namespace NNUE {
Nur in SRC-1.1/src: evaluate_nnue.o.
Nur in SRC-1.1/src: evaluate.o.
Nur in SRC-1.1/src: half_ka_v2_hm.o.
Nur in SRC-1.1/src: main.o.
Nur in SRC-1.1/src: material.o.
diff -ur SRC-1.0a/src/misc.cpp SRC-1.1/src/misc.cpp
--- SRC-1.0a/src/misc.cpp 2022-11-21 19:41:46.000000000 +0100
+++ SRC-1.1/src/misc.cpp 2022-12-21 17:58:32.000000000 +0100
@@ -68,7 +68,7 @@
namespace {
/// Version number or dev.
-const string version = "1.0";
+const string version = "1.1";
/// Our fancy logging facility. The trick here is to replace cin.rdbuf() and
/// cout.rdbuf() with two Tie objects that tie cin and cout to a file stream. We
Nur in SRC-1.1/src: misc.o.
Nur in SRC-1.1/src: movegen.o.
Nur in SRC-1.1/src: movepick.o.
Nur in SRC-1.1/src: pawns.o.
Nur in SRC-1.1/src: polybook.o.
Nur in SRC-1.1/src: position.o.
Nur in SRC-1.1/src: psqt.o.
diff -ur SRC-1.0a/src/search.cpp SRC-1.1/src/search.cpp
--- SRC-1.0a/src/search.cpp 2022-11-21 18:31:06.000000000 +0100
+++ SRC-1.1/src/search.cpp 2022-12-21 18:27:03.000000000 +0100
@@ -787,7 +787,7 @@
// Use static evaluation difference to improve quiet move ordering (~3 Elo)
if (is_ok((ss-1)->currentMove) && !(ss-1)->inCheck && !priorCapture)
{
- int bonus = std::clamp(-19 * int((ss-1)->staticEval + ss->staticEval), -1925, 1925);
+ int bonus = std::clamp(-19 * int((ss-1)->staticEval + ss->staticEval), -1935, 1935);
thisThread->mainHistory[~us][from_to((ss-1)->currentMove)] << bonus;
}
@@ -804,7 +804,7 @@
// If eval is really low check with qsearch if it can exceed alpha, if it can't,
// return a fail low.
if ( depth < 8
- && eval < alpha - 312 - 260 * depth * depth)
+ && eval < alpha - 319 - 260 * depth * depth)
{
value = qsearch<NonPV>(pos, ss, alpha - 1, alpha);
if (value < alpha)
@@ -822,7 +822,7 @@
// The depth condition is important for mate finding.
if ( !ss->ttPv
&& depth < 8
- && eval - futility_margin(depth, improving) - (ss-1)->statScore / 303 >= beta
+ && eval - futility_margin(depth, improving) - (ss-1)->statScore / 298 >= beta
&& eval >= beta
&& eval < 28031) // larger than VALUE_KNOWN_WIN, but smaller than TB wins
return eval;
@@ -833,7 +833,7 @@
&& (ss-1)->statScore < 16979
&& eval >= beta
&& eval >= ss->staticEval
- && ss->staticEval >= beta - 20 * depth - improvement / 13 + 225 + complexity / 24
+ && ss->staticEval >= beta - 20 * depth - improvement / 13 + 220 + complexity / 24
&& !excludedMove
&& pos.non_pawn_material(us)
&& (ss->ply >= thisThread->nmpMinPly || us != thisThread->nmpColor))
@@ -877,7 +877,7 @@
}
}
- probCutBeta = beta + 189 - 49 * improving;
+ probCutBeta = beta + 189 - 51 * improving;
// Step 10. ProbCut (~4 Elo)
// If we have a good enough capture and a reduced search returns a value
@@ -946,7 +946,7 @@
moves_loop: // When in check, search starts here
// Step 12. A small Probcut idea, when we are in check (~0 Elo)
- probCutBeta = beta + 424;
+ probCutBeta = beta + 428;
if ( ss->inCheck
&& !PvNode
&& depth >= 2
@@ -1041,7 +1041,7 @@
&& !PvNode
&& lmrDepth < 7
&& !ss->inCheck
- && ss->staticEval + 204 + 195 * lmrDepth + PieceValue[EG][pos.piece_on(to_sq(move))]
+ && ss->staticEval + 224 + 188 * lmrDepth + PieceValue[EG][pos.piece_on(to_sq(move))]
+ captureHistory[movedPiece][to_sq(move)][type_of(pos.piece_on(to_sq(move)))] / 6 < alpha)
continue;
@@ -1065,7 +1065,7 @@
// Futility pruning: parent node (~9 Elo)
if ( !ss->inCheck
&& lmrDepth < 13
- && ss->staticEval + 120 + 139 * lmrDepth + history / 52 <= alpha)
+ && ss->staticEval + 135 + 133 * lmrDepth + history / 52 <= alpha)
continue;
// Prune moves with negative SEE (~3 Elo)
@@ -1106,7 +1106,7 @@
// Avoid search explosion by limiting the number of double extensions
if ( !PvNode
- && value < singularBeta - 21
+ && value < singularBeta - 32
&& ss->doubleExtensions <= 9)
extension = 2;
}
@@ -1131,7 +1131,7 @@
// Check extensions (~1 Elo)
else if ( givesCheck
&& depth > 9
- && abs(ss->staticEval) > 81)
+ && abs(ss->staticEval) > 84)
extension = 1;
// Quiet ttMove extensions (~0 Elo)
@@ -1204,14 +1204,14 @@
r--;
// Increase reduction if next ply has a lot of fail high
- if ((ss+1)->cutoffCnt > 2 && !PvNode)
+ if ((ss+1)->cutoffCnt > 4 && !capture)
r++;
ss->statScore = 2 * thisThread->mainHistory[us][from_to(move)]
+ (*contHist[0])[movedPiece][to_sq(move)]
+ (*contHist[1])[movedPiece][to_sq(move)]
+ (*contHist[3])[movedPiece][to_sq(move)]
- - 4433;
+ - 4388;
// Decrease/increase reduction for moves with a good/bad history (~30 Elo)
r -= ss->statScore / (16434 - 5880 / depth);
@@ -1229,9 +1229,11 @@
// Adjust full depth search based on LMR results - if result
// was good enough search deeper, if it was bad enough search shallower
const bool doDeeperSearch = value > (alpha + 64 + 11 * (newDepth - d));
+ const bool doEvenDeeperSearch = value > alpha + 582 && ss->doubleExtensions <= 5;
const bool doShallowerSearch = value < bestValue + newDepth;
- newDepth += doDeeperSearch - doShallowerSearch;
+ ss->doubleExtensions = ss->doubleExtensions + doEvenDeeperSearch;
+ newDepth += doDeeperSearch - doShallowerSearch + doEvenDeeperSearch;
if(newDepth>d)
value = -search<NonPV>(pos, ss+1, -(alpha+1), -alpha, newDepth, !cutNode);
Nur in SRC-1.1/src: search.o.
Nur in SRC-1.1/src: tbprobe.o.
Nur in SRC-1.1/src: thread.o.
Nur in SRC-1.1/src: timeman.o.
Nur in SRC-1.1/src: tt.o.
Nur in SRC-1.1/src: tune.o.
Nur in SRC-1.1/src: uci.o.