Komodo 9.4

[lkaufman]

More realistic from human point of view, but also produces better results on very short searches (fixed depth or game in one second plus 10 ms). I'm not sure why the value needs to be higher for real game play.

Higher values of king safety tend to be better for longer time control games. Is 100% like this for Andscacs. Don't know for mobility.

Any idea why this is so?

I have a theory about this:
* large king safety creates large discontinuities in eval (ie. large jumps from one position to the next, just playing a quiet move, involving no capture/promotion).
* large discontinuities hurt your branching factor.

So the tradeoff is:
* more king safety = better eval = more elo in fixed depth testing.
* however, more king safety = worse branching factor = less elo in tc testing.

The tradeoff is either positive or negative depending on the tc. This is normal, because at short tc, hence shallow depth, the extra depth you gain from better BF more than compensates the worse eval. And at long tc, the better eval is more important than the extra depth (which matters less as extra depth gives diminishing returns).

This theory explains a more general empirical truth: any eval component that has large discontinuities (which is inevitable when eval components can have large absolute values) will have this kind of scaling property.

Another example is passed pawn evaluation. It falls exactly in the same bucket as king safety (large values hence large discontinuities). And it also needs longer tc to be tested.

That all sounds sensible, but I don't think it explains the actual data very well. What I observe is that at short fixed depths, the optimal value for king safety rises with depth, while the time taken doesn't vary significantly. There seems to be some reason that higher king safety works well with increased depth, independent of any effect on time taken.

That's weird. Everytime I introduced some eval feature that could be large in absolute value, and introduce large discontinuities, it increased my node count on average. What/how are you measuring exactly ?

Our tester reports nodes; I'm talking about very short searches, just a few ply, and the number of nodes varies by just a fraction of a percent for fairly large swings in weights. I think you can't duplicate this with SF because your pruning at PV nodes makes such short searches worthless.

[Nordlandia]

I choose bishop retreat rather than capture on f8 to retain some attacking possibilities for white.

Overnight game at 180+30 TC.

i7-5960X @3.9GHz
4096 MB per engine were available
6-Men Syzygy
Fritz 15 [Patch 9]
[pgn][Event "EXPERT-PC, Slow 180m+30s"]
[Site "EXPERT-PC"]
[Date "2016.03.20"]
[Round "1"]
[White "Komodo 9.4 64-bit"]
[Black "Stockfish 150316 64 BMI2"]
[Result "1/2-1/2"]
[ECO "A36"]
[Annotator "0.00;-0.01"]
[SetUp "1"]
[FEN "1rbqnrk1/4ppBp/p2p2p1/2p5/2P1PP2/3P2PP/1P1QN1B1/q2NRRK1 w - - 0 17"]
[PlyCount "256"]
[TimeControl "10800+30"]

17. Bc3 e6 {-0.01/34 313} 18. b3 {0.00/36 334} Qa3 {0.63/34 344} (18... Qa3 19.
Nc1 Rxb3 {-0.32/27 Stockfish 6 64}) 19. Nc1 {0.27/35 278} Rb4 {0.88/35 540
(Rxb3)} (19... Rxb3 20. Nxb3 Qxb3 {-0.40/20 Komodo 9.3 64-bit}) 20. Ne3 {
1.01/28 117} Qe7 {1.03/32 651 (f6)} (20... Rxb3 21. Nxb3 Qxb3 {0.31/22 Houdini
4 Pro x64B}) 21. d4 {1.35/30 217 (Nc2)} Rxb3 {1.32/34 514} (21... Qa7 22. f5 {
1.02/26 Stockfish 150316 64 POPCNT}) 22. Nxb3 {1.54/33 131} Qxb3 {1.76/41 1526}
(22... Qxb3 23. Rb1 {1.18/29 Komodo 9.3 64-bit}) 23. dxc5 {2.13/34 363} dxc5 {
1.87/39 656 (Qa4)} (23... dxc5 24. Rb1 Qa4 {0.91/32 Stockfish 5 64 SSE4.2}) 24.
Rb1 {2.19/34 122} Qa4 {1.92/36 75} (24... Qa4 25. Rfd1 Nf6 {0.00/20 Fritz 15
sygyzy}) 25. Qb2 {2.22/34 173 (Rfd1)} f6 {2.42/34 624} 26. Rfd1 {1.92/33 197}
Nd6 {2.49/37 479} (26... Ng7 27. Ng4 Nh5 {1.16/31 Stockfish 5 64 SSE4.2}) 27.
Ra1 {2.09/33 219 (e5)} Nxc4 {1.00/33 252} (27... Qae8 28. Ng4 Nxc4 {0.72/21
Deep Fritz 14 x64}) 28. Rxa4 {2.14/32 225} Nxb2 {1.20/33 75} (28... Nxb2 29.
Bxb2 Rd8 {1.27/16 Deep Rybka 4 x64}) 29. Bxb2 {1.97/34 138} Rd8 {1.08/35 274} (
29... Rd8 30. Rc1 {2.35/26 Komodo 9.4 64-bit}) 30. Rc1 {1.96/36 841} Rd3 {
1.01/36 189} (30... Qb7 31. Rc2 {1.43/27 Stockfish 6 64}) 31. Ra3 {1.78/33 135}
(31. Ra3 Rxa3 {0.97/27 Stockfish 7 64 POPCNT}) 31... Rxa3 {0.95/36 104} (31...
Rxa3 32. Bxa3 Qd8 {0.82/23 Houdini 3 x64}) 32. Bxa3 {1.78/35 188} Bb7 {0.90/36
457 (h5)} (32... Qd8 33. Bxc5 a5 {1.16/22 Fritz 13}) 33. Bxc5 {1.82/33 249} Qd8
{0.87/30 40} (33... Qd8 34. Rb1 h5 {1.20/20 Fritz 11 SE}) 34. Rd1 {1.75/37 412
(Rb1)} Qc7 {0.71/38 938} (34... Qc7 35. Bd4 {1.67/27 Komodo 9.3 64-bit}) 35.
Bd4 {1.80/34 188} Kf7 {0.74/30 38} (35... e5 36. fxe5 fxe5 {0.77/23 Deep Fritz
13}) 36. Kh2 {1.82/36 558 (Rb1)} h5 {0.79/38 197} (36... h5 37. Rb1 {1.61/27
Komodo 9.3 64-bit}) 37. h4 {1.77/39 384 (Rb1)} a5 {0.79/36 101} (37... a5 38.
Bb2 a4 {1.22/19 Deep Rybka 4 x64}) 38. Rb1 {1.71/41 438} Ba8 {0.79/38 36} (
38... Ba8 39. Bb2 f5 {0.84/21 Deep Fritz 14 x64}) 39. Bb2 {1.80/40 285} Qc5 {
0.78/41 113} (39... f5 40. Rc1 Qd8 {1.15/33 Stockfish 5 64 SSE4.2}) 40. Rc1 {
1.80/41 136} Qxe3 {0.78/43 42} (40... Qxe3 41. Rc7+ Ke8 {1.85/33 Stockfish 4 64
}) 41. Rc7+ {1.55/41 103} Ke8 {0.78/46 169 (Kf8)} (41... Ke8 42. Rc8+ Kf7 {
1.03/37 Stockfish 5 64 SSE4.2}) 42. Rc8+ {1.48/41 110} Kf7 {0.78/43 77 (Ke7)} (
42... Kf7 43. Rxa8 Qc5 {1.53/26 Fritz 15}) 43. Rxa8 {2.02/38 124} Qc5 {0.78/42
211} (43... Qc5 44. Rb8 e5 {0.62/24 Houdini 3 Pro x64}) 44. Rh8 {1.94/42 470}
Qb5 {0.78/42 126 (e5)} 45. Bd4 {1.87/40 267 (Ba1)} e5 {0.78/45 90} (45... e5
46. fxe5 fxe5 {0.84/34 Stockfish 100316 64 POPCNT}) 46. fxe5 {1.83/43 95} fxe5
{0.78/46 65} (46... fxe5 47. Bc3 a4 {0.94/26 Deep Fritz 14 x64}) 47. Be3 {
1.81/47 438} Qb4 {0.78/46 78 (a4)} 48. Rh7+ {1.80/45 180 (Ra8)} Kg8 {0.78/48
150} (48... Kg8 49. Rc7 a4 {0.50/29 Houdini 4 Pro x64}) 49. Ra7 {1.80/47 183}
a4 {0.78/50 37} (49... a4 50. Bh6 Qd4 {0.83/27 Deep Fritz 14 x64}) 50. Ra6 {
1.80/47 204 (Ra8+)} Kf7 {0.78/45 70} (50... Kf7 51. Ra7+ Kg8 {1.36/29 Fritz 15}
) 51. Bg5 {1.78/45 202 (Bh6)} Qb3 {0.40/45 76 (Qd4)} (51... Qb3 52. Ra7+ Ke8 {
0.21/28 Houdini 4 Pro x64B}) 52. Ra7+ {1.74/47 197} Ke8 {0.40/47 196} (52...
Ke8 53. Bf6 a3 {1.07/29 Fritz 15}) 53. Re7+ {1.72/47 165 (Bf6)} Kf8 {0.40/48 87
} (53... Kf8 54. Ra7 Ke8 {1.51/37 Komodo 9.3 64-bit}) 54. Ra7 {1.71/48 149
(Rxe5)} Ke8 {0.00/69 69} (54... Ke8 55. Bf6 a3 {1.12/19 Fritz 6}) 55. Re7+ {
1.71/46 74} Kf8 {0.00/106 89} 56. Rxe5 {1.70/47 77} a3 {0.40/50 244} (56... a3
57. Ra5 a2 {0.68/29 Deep Fritz 14 x64}) 57. Ra5 {1.71/45 79} a2 {0.40/48 21} (
57... a2 58. Bf6 Qc2 {0.33/41 SugaR PrO v1.3 64 POPCNT}) 58. Bf6 {1.70/44 95}
Kf7 {0.40/50 340 (Kg8)} (58... Qb1 59. Be5 Ke8 {0.69/25 Deep Fritz 14 x64}) 59.
Be5 {1.70/45 97 (Bd4)} Qb1 {0.40/48 137 (Qc2)} (59... Qb1 60. Ra6 Qb5 {1.05/36
Stockfish 4 64}) 60. Ra7+ {1.69/45 117 (Bc3)} Ke8 {0.40/48 111 (Ke6)} (60...
Ke8 61. Bd4 Qd1 {1.82/37 Komodo 9.4 64-bit}) 61. Bh8 {1.69/45 95 (Bg7)} Qb3 {
0.40/49 50 (Qc2)} (61... Qc2 62. Ra6 Kf7 {1.54/38 Komodo 9.3 64-bit full
Syzygy on SSD}) 62. Ba1 {1.69/43 89 (Ra6)} Qc2 {0.40/49 120 (Qc4)} (62... Qb1
63. Bd4 Qd1 {1.40/35 Komodo 9.3 64-bit}) 63. Ra6 {1.68/45 337 (Bd4)} Qb1 {
0.40/46 70} (63... Kf7 64. Ra7+ Ke8 {1.44/35 Komodo 9.3 64-bit}) 64. Bh8 {
1.68/46 158 (Bf6)} Kf7 {0.40/51 176} (64... Kf7 65. Bc3 Qb3 {0.61/26 Deep
Fritz 14 x64}) 65. Ra7+ {1.69/46 272 (Be5)} Kg8 {0.40/52 61 (Ke8)} (65... Kf8
66. Bg7+ Ke8 {1.77/29 Komodo 8 64-bit}) 66. Be5 {1.67/46 117 (Bd4)} Qb5 {
0.40/48 166} (66... Qb5 67. Ba1 Qb1 {0.82/26 Deep Rybka 4 w32}) 67. Ra8+ {
1.66/45 127} Kf7 {0.40/45 14} (67... Kf7 68. Bh8 Qb1 {1.32/33 Komodo 9.1 32-bit
}) 68. Bc3 {1.66/40 56 (Bd4)} Qd3 {0.40/48 84 (Qb1)} (68... Qc6 69. Ra7+ Ke8 {
1.79/30 Komodo 8 64-bit}) 69. Ra7+ {1.63/45 134} Ke8 {0.40/50 17 (Kg8)} (69...
Ke8 70. Bb2 Qb3 {0.23/40 Stockfish 7 64}) 70. Bb2 {1.63/44 104} Qc4 {0.40/52
43 (Qb3)} (70... Qb1 71. Ra8+ Kf7 {0.00/26 Fritz 13 SE}) 71. Ra3 {1.63/43 129
(Ra5)} Kf7 {0.40/49 57 (Ke7)} (71... Ke7 72. Ra7+ Ke8 {0.66/26 Deep Fritz 14
x64}) 72. Bh8 {1.63/44 73 (Ra8)} Qe2 {0.40/47 53} (72... Qe2 73. Ra7+ Ke8 {
0.47/36 Stockfish 7 64}) 73. Ra7+ {1.63/41 50} Ke8 {0.40/45 100 (Kf8)} (73...
Ke8 74. Ba1 Qc2 {0.65/26 Deep Fritz 14 x64}) 74. Ra8+ {1.62/41 57 (Be5)} Kd7 {
0.40/49 161 (Kf7)} (74... Kf7 75. Ra7+ Ke6 {1.88/36 Komodo 9.4 64-bit}) 75. Be5
{1.62/40 134 (Bg7)} Ke7 {0.40/46 42 (Ke6)} (75... Ke6 76. Ba1 Ke7 {0.00/25
Deep Fritz 13}) 76. Ra7+ {1.60/42 415 (Ra3)} Kf8 {0.40/49 83 (Ke6)} (76... Ke8
77. Ra8+ Kf7 {0.62/27 Deep Fritz 14 x64}) 77. Bd4 {1.60/39 65 (Bg7+)} Ke8 {
0.40/52 34} (77... Qc4 78. Bf6 Qc2 {1.82/38 Komodo 9.4 64-bit}) 78. Ra8+ {
1.57/38 106 (Ra5)} Ke7 {0.40/52 76 (Kd7)} (78... Kf7 79. Ra7+ Kf8 {1.52/31
Komodo 9.3 64-bit}) 79. Be5 {1.57/39 88 (Ra3)} Ke6 {0.40/51 55} (79... Ke6 80.
Bc3 Ke7 {0.33/37 Stockfish 7 64}) 80. Bd4 {1.58/37 122 (Bg7)} Ke7 {0.00/51 40
(Kf7)} 81. Ra4 {1.57/42 89} Qd2 {0.40/47 124 (Qc2)} (81... Kd6 82. e5+ Ke6 {
0.25/23 Fritz 11 SE}) 82. Ra7+ {1.56/41 53 (Be5)} Ke8 {0.40/48 126} 83. Be5 {
1.56/43 57} Qe2 {0.40/46 10 (Qf2)} (83... Qe2 84. Bh8 Qc4 {0.30/41 Stockfish 6
64}) 84. Ra5 {1.56/42 54 (Ra3)} Kd7 {0.40/48 41 (Kf7)} (84... Kd7 85. Ra7+ Ke8
{0.92/25 Deep Fritz 14 x64}) 85. Ra8 {1.56/41 49 (Bd4)} Ke6 {0.00/55 32} 86.
Bc3 {1.54/40 86} Kf7 {0.40/50 309 (Ke7)} (86... Qc2 87. Bd4 Qd2 {1.36/32
Komodo 9.3 32-bit}) 87. Ra7+ {1.56/39 115 (Ra5)} Ke8 {0.39/44 25} (87... Ke8
88. Ra8+ Kf7 {0.57/26 Deep Fritz 14 x64}) 88. Bd4 {1.49/40 58 (Bf6)} Qd2 {
0.00/55 25 (Qd1)} (88... Kf8 89. Bc3 Ke8 {0.24/33 Stockfish 7 64}) 89. Bf6 {
1.49/40 44} Qe2 {0.39/45 58 (Qc2)} (89... Qc2 90. Bd4 Qd3 {0.22/29 Houdini 4
x64}) 90. Be5 {1.46/39 61 (Ra4)} Kf8 {0.00/59 91 (Qc4)} (90... Qb5 91. Ra8+ Kf7
{0.75/27 Deep Fritz 14 x64}) 91. Bf6 {1.32/40 68} Ke8 {0.00/57 28 (Qf2)} (91...
Qf2 92. Ra8+ Kf7 {0.70/30 Deep Fritz 14 x64}) 92. Bh8 {1.31/38 31} Kd8 {
0.39/44 41 (Qc2)} (92... Qc2 93. Bg7 Qd2 {1.86/32 Komodo 9.02 64-bit}) 93. Bd4
{1.08/36 88 (Bg7)} Ke8 {0.00/57 19 (Qd2)} (93... Qd2 94. Ba1 Ke8 {1.41/32
Komodo 9.3 64-bit}) 94. Bf6 {0.99/36 42} Kf8 {0.39/44 129 (Qf2)} (94... Qf2 {
1.37/34 Komodo 9.1 64-bit}) 95. Ra3 {0.91/34 22 (Bg7+)} Kf7 {0.32/36 20 (Qc2)}
(95... Ke8 96. Bd4 Qd2 {1.54/35 Komodo 9.3 64-bit full Syzygy on SSD}) 96. Bd4
{0.89/34 41 (Bh8)} Ke7 {0.39/35 43 (Qc2)} (96... Qd1 97. Ra7+ Ke8 {0.27/31
Houdini 3 Pro x64}) 97. Ra4 {0.83/36 48 (Bg7)} Qd2 {0.00/57 23 (Kf7)} 98. Bh8 {
0.81/37 35} Qe2 {0.00/36 28} (98... Qe2 99. Ba1 {0.84/26 Deep Fritz 14 x64})
99. Ba1 {0.77/38 28 (Be5)} (99. Ra7+ Kf8 100. Bg7+ {0.64/27 Deep Fritz 14 x64})
99... Ke8 {0.00/45 20} (99... Kf7 100. Ra7+ Kf8 {0.73/27 Deep Fritz 14 x64})
100. e5 {0.77/37 20 (Be5)} (100. Ra8+ Ke7 101. Ra7+ {0.33/26 Houdini 4 x64B})
100... Kd7 {0.00/37 16 (Ke7)} (100... Ke7 101. Bc3 Kd7 {0.98/28 Fritz 15}) 101.
Ra7+ {0.75/39 37 (Bc3)} (101. Ra7+ Ke6 {1.25/27 Komodo 9.1 32-bit}) 101... Ke8
{0.00/40 28} (101... Ke6 102. Bd4 Qc2 {0.58/26 Deep Fritz 14 x64}) 102. Ra8+ {
0.57/36 141 (Bc3)} Ke7 {0.00/39 19} (102... Kf7 103. Ra4 Qd2 {1.05/26 Fritz 15
sygyzy}) 103. Ra7+ {0.55/37 49 (Bc3)} Ke8 {0.00/60 16 (Ke6)} (103... Ke6 104.
Ra4 {0.82/29 Komodo 9.3 64-bit}) 104. e6 {0.51/35 55} Qxe6 {0.00/42 25} 105.
Ra5 {0.42/36 33 (Bb2)} Kf8 {0.00/49 26 (Qb3)} (105... Qc4 106. Bb2 Qb3 {
0.00/41 Stockfish 210216 64}) 106. Ra8+ {0.40/37 56} Ke7 {0.00/53 89 (Kf7)} (
106... Kf7 107. Ra7+ Kg8 {0.00/41 Stockfish 7 64}) 107. Ra4 {0.37/37 17 (Ra7+)}
(107. Ra7+ Ke8 108. Ra5 {0.97/28 Komodo 9.02 64-bit}) 107... Kd8 {0.00/39 19
(Qb3)} (107... Qb3 108. Ra6 Qc2 {0.61/32 Komodo 9.3 64-bit}) 108. Rd4+ {
0.36/34 17 (Ra8+)} (108. Rd4+ Kc7 109. Rd2 {0.12/27 Houdini 3 Pro x64}) 108...
Kc7 {0.00/39 25} (108... Kc7 109. Rd2 g5 {0.36/24 Deep Fritz 14 x64}) 109. Re4
{0.35/36 20 (Rd2)} (109. Rd2 g5 110. Bc3 {0.45/23 Deep Fritz 14 x64}) 109...
Qa6 {0.00/46 15} (109... Qb3 110. Re7+ Kd6 {0.00/38 Stockfish 7 64}) 110. Be5+
{0.34/33 28} (110. Rb4 Qa5 111. Rb2 {0.36/25 Deep Fritz 14 x64}) 110... Kb6 {
0.00/49 17 (Kd7)} (110... Kc8 111. Bh3+ Kd8 {0.00/41 Stockfish 150316 64 POPCNT
}) 111. Bd4+ {0.29/35 68} Kc7 {0.00/56 6} (111... Kc7 112. Ba1 Kd6 {0.43/32
Komodo 9.3 64-bit}) 112. Ba1 {0.28/33 8 (Be5+)} (112. Ba1 Kd6 {0.69/34 Komodo
9.4 64-bit}) 112... Kd7 {0.00/43 17} (112... Qd3 113. Re6 Kd8 {0.00/29 Houdini
4 Pro x64}) 113. Rf4 {0.21/33 52 (Bh3+)} Qe2 {0.00/43 30 (Qb5)} (113... Qe2
114. Rf7+ Ke8 {0.00/39 Stockfish 5 64 SSE4.2}) 114. Rd4+ {0.20/37 38} Kc7 {
0.00/47 18 (Kc8)} (114... Kc7 115. Bc3 Kb6 {0.00/40 Stockfish 7 64}) 115. Ra4 {
0.20/35 10} (115. Bc3 Qa6 116. Rd1 {0.00/27 Deep Fritz 14 x64}) 115... Qe6 {
0.00/47 30 (Kd7)} (115... Qf2 116. Ra6 Qe2 {0.59/31 Komodo 9.3 64-bit}) 116.
Bb2 {0.20/38 17 (Ra7+)} (116. Ra7+ Kd8 117. Bb2 {0.00/40 Stockfish 7 64})
116... Qb3 {0.00/42 22 (Kd8)} (116... Kb6 117. Bf1 Qb3 {0.00/28 Deep Fritz 14
x64}) 117. Ra7+ {0.19/35 17} (117. Ra7+ Kd6 118. Bf6 {0.57/34 Komodo 9.3 64-bit
}) 117... Kc8 {0.00/44 16 (Kd6)} (117... Kd6 118. Bd4 Qc4 {0.00/29 Deep Fritz
14 x64}) 118. Bf6 {0.19/37 12} (118. Bd4 Qc2 119. Ra8+ {0.00/42 Stockfish
210216 64}) 118... Qb1 {0.00/41 22 (Qc2)} (118... Qc2 119. Rg7 Kb8 {0.00/42
Stockfish 7 64}) 119. Bh3+ {0.18/39 22 (Ra8+)} (119. Be5 Qc2 120. Ra8+ {
0.95/30 Komodo 9 64-bit}) 119... Kb8 {0.00/45 21} (119... Kb8 120. Rg7 Qc2+ {
0.00/27 Deep Fritz 14 x64}) 120. Ra5 {0.17/36 19 (Ra3)} (120. Ra6 Kc7 121. Be5+
{0.00/39 Stockfish 6 64}) 120... Kb7 {0.00/44 22 (Qc2+)} (120... Qc2+ 121. Bg2
Kc7 {0.59/35 Komodo 9.1 64-bit}) 121. Bg2+ {0.15/39 88} Kc8 {0.00/45 18 (Kc7)}
(121... Kc7 122. Ra6 Kd7 {0.00/40 Stockfish 5 64 SSE4.2}) 122. Ra7 {0.13/38 70
(Ra8+)} (122. Bf3 Qf1 123. Be4 Qf2+) 122... Qc2 {0.00/48 24} 123. Bh8 {0.11/34
9 (Ra8+)} (123. Rg7 Kb8 124. Be5+ {0.00/43 Stockfish 7 64}) 123... Qf2 {
0.00/39 22 (Kd8)} (123... Kd8 124. Bd4 Qe2 {0.00/27 Houdini 3 x64}) 124. Ra3 {
0.09/36 15 (Ra6)} (124. Ra6 Qe2 125. Rxg6 {0.13/24 Fritz 11 SE}) 124... Kd7 {
0.00/45 30} (124... Qe2 125. Kg1 Kd7 {0.00/40 Stockfish 6 64}) 125. Ra4 {
0.09/37 37 (Ra6)} (125. Ra6 Qc2 126. Ra7+ {0.00/27 Deep Fritz 14 x64}) 125...
Kc7 {0.00/39 22 (Qc2)} (125... Qe2 126. Bd4 Kd6 {0.00/31 Deep Fritz 14 x64})
126. Rf4 {0.09/37 25 (Ra8)} (126. Ra6 Qe2 127. Rc6+ {0.25/39 Stockfish 7 64})
126... Qc5 {0.00/39 24} (126... Qe1 127. Rf7+ Kd8 {0.00/40 Stockfish 150316 64
POPCNT}) 127. Ra4 {0.08/36 34 (Rf7+)} (127. Rf6 Qa5 128. Rxg6 {0.11/25 Deep
Fritz 14}) 127... Qc2 {0.00/38 49} (127... Qc2 128. Ra6 Kd7 {0.00/40 Stockfish
160715 64 POPCNT}) 128. Ra8 {0.08/34 16} (128. Ra6 Kd7 129. Bd4 {0.00/38
Stockfish 160715 64 POPCNT}) 128... Kb6 {0.00/37 28 (Kd6)} (128... Kb6 129.
Bd4+ Kb5 {0.00/26 Deep Rybka 4 SSE42 x64}) 129. Be5 {0.08/35 21 (Kg1)} (129.
Kg1 Qe2 130. Ra3 {0.00/34 Stockfish 5 64 SSE4.2}) 129... Kb5 {0.00/50 25 (Qe2)}
(129... Kb5 130. Rb8+ Ka4 {0.16/30 Komodo 9.3 64-bit}) 130. Rb8+ {0.08/35 26} (
130. Rb8+ Ka4 131. Ra8+ {0.00/46 Stockfish 7 64 POPCNT}) 130... Ka4 {0.00/54 9
(Ka5)} (130... Ka4 131. Ra8+ Kb3 {0.00/52 Stockfish 7 64}) 131. Rf8 {0.07/35
24 (Ra8+)} (131. Ra8+ Kb5 132. Rb8+ {0.00/33 Deep Fritz 14 x64}) 131... Qc5 {
0.00/44 22 (Kb5)} (131... Qc5 132. Ra8+ Kb3 {0.00/37 Stockfish 5 64 SSE4.2})
132. Rf4+ {0.07/36 15 (Ra8+)} (132. Ra8+ Kb3 133. Rb8+ {0.00/41 Stockfish
150316 64 POPCNT}) 132... Kb3 {0.00/43 21} (132... Kb5 133. Bg7 Qe7 {0.00/35
Komodo 9.3 64-bit}) 133. Bf6 {0.04/38 61 (Rf3+)} Qc2 {0.00/43 21 (Qd6)} (133...
Qc2 134. Kh1 Qd1+ {0.00/36 Stockfish 100316 64 POPCNT}) 134. Rf3+ {0.00/40 18}
(134. Kh1 Qd1+ 135. Rf1 {0.00/36 Stockfish 150316 64 POPCNT}) 134... Ka4 {
0.00/43 24 (Kb4)} (134... Kb4 135. Ba1 Qd1 {0.00/29 Deep Fritz 14 x64}) 135.
Rf1 {0.00/43 31 (Rf4+)} (135. Bg7 Qc7 136. Bd4 {0.07/31 Komodo 9.3 64-bit})
135... Qe2 {0.00/44 31 (Qd2)} (135... Kb5 136. Kg1 Qd3 {0.00/30 Houdini 3 x64})
136. Bh8 {0.00/44 64 (Rf3)} Kb3 {0.00/45 22 (Qc2)} (136... Qd2 137. Rf8 Qd1 {
-0.25/21 Fritz 11 SE}) 137. Rf3+ {0.00/42 14 (Bf6)} (137. Rf8 Qd1 138. Rf3+ {
0.10/28 Komodo 9.3 32-bit}) 137... Kc2 {0.00/44 22 (Kc4)} (137... Kb4 138. Rf8
Qc2 {-0.03/16 Fritz 6}) 138. Rc3+ {0.13/32 24 (Rf4)} Kb1 {0.00/42 34 (Kd2)} (
138... Kd2 139. Rc6 Qe8 {0.00/36 Stockfish 100316 64 POPCNT}) 139. Ra3 {
0.13/33 21 (Rb3+)} (139. Rb3+ Kc1 140. Rb6 {0.00/26 Deep Fritz 14}) 139... Qe8
{0.00/42 24 (Qc4)} (139... Qc2 140. Re3 Qd2 {0.00/26 Deep Fritz 14 x64}) 140.
Bf6 {0.00/40 20} (140. Ra8 Qe2 141. Rb8+ {0.00/26 Deep Fritz 14 x64}) 140...
Qe2 {0.00/46 24} (140... Qe2 141. Ra7 Qc4 {0.00/29 Houdini 3 x64}) 141. Bh8 {
0.00/47 50 (Rb3+)} Kc1 {0.00/46 55 (Qe8)} 142. Rb3 {0.00/43 33 (Rc3+)} (142.
Rc3+ Kd1 143. Rc6 {0.00/37 Stockfish 7 64}) 142... Qe6 {0.00/42 23 (Qd2)} (
142... Qd2 143. Rc3+ Kd1 {0.00/26 Houdini 4 Pro x64}) 143. Rf3 {0.00/41 16
(Rc3+)} (143. Rc3+ Kd1 144. Rc6 {0.00/31 Houdini 4 Pro x64B}) 143... Kd1 {
0.00/44 22 (Kd2)} (143... Kd2 144. Rf2+ Ke3 {0.33/33 Komodo 9.3 64-bit}) 144.
Bd4 {0.00/43 24 (Rf2)} (144. Rf6 Qg8 145. Rd6+ {0.00/29 Houdini 4 x64}) 144...
Qd6 {0.00/42 25 (Qe2) Draw accepted} (144... Qe2 145. Bg7 Kc2 {0.34/34 Komodo
9.3 64-bit}) 1/2-1/2[/pgn]

[Leto]

The Readme file is very well done as background material and explanation.
For analysis, it is unclear why or when one would ever use Contempt for White, since this parameter only weakens the analysis.
Also, my default value is "0", not "10".

Contempt for White would be used for example if yiou are playing a correspondence game as White against a weaker player, and you want to avoid draws. So maybe raise the default value of 0 (10 is default for regular contempt) to 10 or 20 to analyze with White trying to avoid draws. If you are Black just use a negative number for Contempt for White.

On my 64-bit Komodo, default for Contempt is "0". Is the correct default value "10"?
"So maybe raise the default value of 0 (10 is default for regular contempt)"
What has a default value of "0"?
When I check the box "Contempt for White", I don't see a separate box to enter a number for "Contempt for White".
My understanding of checking the "Cfw" box is that all it does is allow contempt values to be used during analysis.

On Komodo 9.4 the default "play" contempt is 10. In analyis mode Komodo always uses contempt 0 unless you have the "Contempt for White" enabled, which then uses the "play" contempt in analysis mode.

[cma6]

"On Komodo 9.4 the default "play" contempt is 10. In analyis mode Komodo always uses contempt 0 unless you have the "Contempt for White" enabled, which then uses the "play" contempt in analysis mode."

Leto,
Thanks much for finally explaining the mechanism of the relationship between Contempt and Cfw.

Your explanation could profitably be added to the Readme file.

[cma6]

Higher values of king safety tend to be better for longer time control games. Is 100% like this for Andscacs. Don't know for mobility.[/quote]

Another example is passed pawn evaluation. It falls exactly in the same bucket as king safety (large values hence large discontinuities). And it also needs longer tc to be tested.[/quote]

Therefore to experts like LK and Lucas, I pose this practical question:
Assuming one is using longer tc (3-4 minutes/move analysis) and assuming there are passed P and/or K-safety issues on the board, how should one change Dynamism from 100 default in such positions?

[hgm]

I have a theory about this:
* large king safety creates large discontinuities in eval (ie. large jumps from one position to the next, just playing a quiet move, involving no capture/promotion).
* large discontinuities hurt your branching factor.

The branching factor does not suffer so much from large evaluation terms, as well from gross misevaluations. Captures and promotions also give large evaluation swings from on position to the next. Reducing the weight of material in the evaluation is not the answer; quiescence search is.

Evaluation terms that need to be large and can be hugely changed by a single move should be verified by searching the moves that change them for the good, to make sure they are not illusory. Then you will have a stable search and a good branching factor.

Making a tradeoff between two evils is usually not the best thing one can do.

[lkaufman]

I have a theory about this:
* large king safety creates large discontinuities in eval (ie. large jumps from one position to the next, just playing a quiet move, involving no capture/promotion).
* large discontinuities hurt your branching factor.

The branching factor does not suffer so much from large evaluation terms, as well from gross misevaluations. Captures and promotions also give large evaluation swings from on position to the next. Reducing the weight of material in the evaluation is not the answer; quiescence search is.

Evaluation terms that need to be large and can be hugely changed by a single move should be verified by searching the moves that change them for the good, to make sure they are not illusory. Then you will have a stable search and a good branching factor.

Making a tradeoff between two evils is usually not the best thing one can do.

Yes, we could fix this position by including castling in qsearch, although it might hurt elo overall, and if we took away castling rights from both sides in this position that would not help. But the real question is: Why should an eval that is so clearly wrong to any strong chessplayer test so well in engine play? Simply "threatening" a couple random pointless checks and "threatening" to lose your queen cannot be worth two pawns.

[hgm]

Proably because if it would happen at the end of a branch of a deeper search, and it can be easily cured (e.g. by Nc6), it could probably be cured pre-emptively by that same move. So the disadvantaged party would just stear the search such the positions like this cannot be forced. That might not be entirely without cost, eval-wise, but the cost will likely be much smaller than the evaluation in the particular position.

I am not sure that including castling would hurt overall Elo. For most of the game you cannot castle. And you would only need to search it in QS when King Safety is bad, so that it s likely there is something to gain by the castling. (Just like you only consider good captures in QS.)

The same could be said for non-capture Pawn pushes that create a passer.

[lkaufman]

Proably because if it would happen at the end of a branch of a deeper search, and it can be easily cured (e.g. by Nc6), it could probably be cured pre-emptively by that same move. So the disadvantaged party would just stear the search such the positions like this cannot be forced. That might not be entirely without cost, eval-wise, but the cost will likely be much smaller than the evaluation in the particular position.

I am not sure that including castling would hurt overall Elo. For most of the game you cannot castle. And you would only need to search it in QS when King Safety is bad, so that it s likely there is something to gain by the castling. (Just like you only consider good captures in QS.)

The same could be said for non-capture Pawn pushes that create a passer.

These ideas are reasonable, but even without castling the move ..Kf8 avoids the two safe checks; the problem is much more general than just castling option.

[mjlef]

"On Komodo 9.4 the default "play" contempt is 10. In analyis mode Komodo always uses contempt 0 unless you have the "Contempt for White" enabled, which then uses the "play" contempt in analysis mode."

Leto,
Thanks much for finally explaining the mechanism of the relationship between Contempt and Cfw.

Your explanation could profitably be added to the Readme file.

It is in the README file. Here is the section:

The “Contempt for White” setting forces Komodo to always use Contempt for White’s point of view, instead of from whatever side Komodo is searching or analyzing now. A positive score will discourage White from accepting a draw or trading pieces. A negative score will encourage Komodo to trade pieces and seek a draw. For black, the opposite sign should be used (for example, a negative score will discourage black from exchanging pieces). “Contempt for White” also prevents Komodo from automatically setting Contempt to zero when running infinite analysis. This helps make analysis consistent when switching sides and exploring various lines and lets you include a non-zero Contempt in your analysis. Note when playing against the computer, and you wish to use a non-zero Contempt, either turn off “Contempt for White” so that Contempt will apply to the Computer’s side, or you can use the above description to set an appropriate Contempt for the side that Komodo is playing.