Author: Will Singleton
Date: 18:25:07 08/27/02
Go up one level in this thread
On August 27, 2002 at 20:07:55, scott farrell wrote:
>Guys,
>
>I am playing around with IID.
>
>Currently its implemented to do an IID search only where:
>- its a PV node
>- there is no best move already in the hash table
>- (alpha != beta - 1) (this is from R.Hyatt as we both use PVS)
>- its not a qsearch
>- not a null move search
>
>I am trying to see what using IDD at all nodes where (alpha != beta - 1) does
>(other than just burn more nodes). A few things I found:
>- it does go faster
>- it seems to 'cover up' refutations until the next ply or so. Which seems in
>line with playing around with PVS searches. I am judging this on a few test
>positions, plus wac scores at 1 sec/position.
>
>What I want to know is there anything inherently risky with IDD at PVS nodes -
>or is it more like dumb luck that this happens (ie. interaction with something
>else like null moves, or extensions or qsearch or something).
>
>The test position below shows the new way takes until depth 9 to solve with
>B6D7, and the standard IDD (not used at PVS nodes) finds it at depth 7. The
>branching factor (bf) and average bf (avbf) are lower on IDD at all nodes, the
>beta cuttoff on first move (bc%) is always higher, and the PV on first move(pv%)
>is always higher also.
>
>Here is one test position, and my analysis both ways, and crafty's analysis at
>the end:
>
>7k/pp3p1p/1np5/5N2/1P2P1NR/7P/3r2PK/2r5/ b - - 0 50
>
>this is my analysis WITH IDD on PVS nodes as well (modified to
>alpha=alpha-pawn/2 and research at alpha=alpha-pawn)
>
>, D4.0(3): 4.34 : C1C2 , 0.44/99999.0 secs - NPS: 15827 AvCut: 1.3
>itdeep:709/518, ks:6/-50, pp:0(0)/0(0), mateN: 1, mateNQ: 1, threatN: 1/1, N:
>5190, nmN: 1221, nmNq: 45, qN: 508, cHit: 97%, cqHit: 0%, cUse%: 102, bc%:84,
>pv%:55, pvs%:90,pc%:83,pv:C1C2 G4E3 C2A2 E3C2
>, D5.0(3): 5.44 : C1C2 , 0.55/99999.0 secs - NPS: 26044 , bf:2.06, avbf:2.06
>AvCut: 1.5 itdeep:1208/786, ks:6/-50, pp:0(0)/0(0), mateN: 1, mateNQ: 2,
>threatN: 1/2, N: 9889, nmN: 3427, nmNq: 169, qN: 839, cHit: 90%, cqHit: 2%,
>cUse%: 92, bc%:93, pv%:65, pvs%:95,pc%:66,pv:C1C2 H2G3 C2B2 G3F4 B2B4
>, D6.0(5): 4.07 : C1C2 , 0.77/99999.0 secs - NPS: 48717 , bf:2.62, avbf:2.34
>AvCut: 0.5 itdeep:2489/1705, ks:6/-50, pp:0(0)/0(0), mateN: 1, mateNQ: 2,
>threatN: 6/8, N: 23059, nmN: 12629, nmNq: 414, qN: 1410, cHit: 81%, cqHit: 1%,
>cUse%: 82, bc%:91, pv%:70, pvs%:96,pc%:96,pv:C1C2 G4E3 C2B2 E4E5 B2A2 E3C4
>, D7.0(5): 4.69 : C1C2 , 1.7/99999.0 secs - NPS: 75524 , bf:3.42, avbf:2.7
>AvCut: 0.7 itdeep:7605/4497, ks:6/-50, pp:0(0)/0(0), mateN: 38, mateNQ: 41,
>threatN: 65/73, N: 75246, nmN: 46335, nmNq: 2902, qN: 3907, cHit: 88%, cqHit:
>3%, cUse%: 91, bc%:91, pv%:53, pvs%:97,pc%:83,pv:C1C2 G4E3 C2B2 E4E5 F7F6 E5F6
>B2B4
>, D8.0(4): 3.9 : C1C2 , 3.57/99999.0 secs - NPS: 87529 , bf:2.43, avbf:2.63
>AvCut: 0.6 itdeep:16294/10245, ks:6/-50, pp:0(0)/0(0), mateN: 69, mateNQ: 72,
>threatN: 82/155, N: 152362, nmN: 143325, nmNq: 8053, qN: 8738, cHit: 136%,
>cqHit: 3%, cUse%: 142, bc%:87, pv%:63, pvs%:95,pc%:93,pv:C1C2 G4E3 C2C3 H4G4
>B6C4 G4G7 C4E3 F5E3
>, D9.0(5): 4.54 : B6D7 , 12.41/99999.0 secs - NPS: 94122 , bf:3.74, avbf:2.85
>AvCut: 0.8 itdeep:63875/39147, ks:-94/-50, pp:0(0)/0(0), mateN: 617, mateNQ:
>456, threatN: 1104/1259, N: 611739, nmN: 469479, nmNq: 37541, qN: 49297, cHit:
>104%, cqHit: 5%, cUse%: 112, bc%:86, pv%:63, pvs%:95,pc%:85,pcc:2,pv:B6D7 F5H6
>F7F5 E4F5 C1E1 F5F6 E1E2 H6F7 H8G8
>, D10.0(5): 3.85 : B6D7 , 22.46/99999.0 secs - NPS: 96397 , bf:1.85, avbf:2.69
>AvCut: 0.8 itdeep:108994/69136, ks:-94/-50, pp:0(0)/0(0), mateN: 1147, mateNQ:
>583, threatN: 1124/2383, N: 965575, nmN: 1049738, nmNq: 70242, qN: 79514, cHit:
>124%, cqHit: 5%, cUse%: 132, bc%:79, pv%:51, pvs%:93,pc%:93,pcc:15,pv:B6D7 F5H6
>F7F6 H6F7
>, D11.0(6): 4.14 : B6D7 , 45.7/99999.0 secs - NPS: 99338 , bf:2.1, avbf:2.6
>AvCut: 0.9 itdeep:232115/140136, ks:-94/-50, pp:0(0)/0(0), mateN: 3492, mateNQ:
>1537, threatN: 3679/6062, N: 1995167, nmN: 2213245, nmNq: 159329, qN: 171992,
>cHit: 106%, cqHit: 6%, cUse%: 114, bc%:81, pv%:60, pvs%:94,pc%:93,pcc:84,pv:B6D7
>H4H6 C1C2 G4E3 C2B2 H6D6 D7E5
>, D12.0(6): 3.99 : B6D7 , 157.2/99999.0 secs - NPS: 95890 , bf:3.32, avbf:2.69
>AvCut: 0.8 itdeep:769624/426178, ks:-94/-50, pp:0(0)/0(0), mateN: 6731, mateNQ:
>3315, threatN: 9844/15906, N: 5544572, nmN: 8396397, nmNq: 641691, qN: 491247,
>cHit: 91%, cqHit: 3%, cUse%: 95, bc%:81, pv%:62, pvs%:93,pc%:95,pcc:441,pv:B6D7
>H4H6 C1E1
>
>
>This is WITHOUT IDD at PVS
>
>, D5.0(3): 5.44 : C1C2 , 0.55/99999.0 secs - NPS: 23151 , bf:2.1, avbf:2.1
>AvCut: 1.5 itdeep:120/82, ks:6/-50, pp:0(0)/0(0), mateNQ: 4, threatN: 4/5, N:
>8428, nmN: 3606, nmNq: 179, qN: 520, cHit: 52%, cqHit: 2%, cUse%: 54, bc%:92,
>pv%:54, pvs%:94,pc%:65,pv:C1C2 H2G3 C2B2 G3F4 B2B4
>, D6.0(5): 4.07 : C1C2 , 0.77/99999.0 secs - NPS: 42208 , bf:2.55, avbf:2.33
>AvCut: 0.6 itdeep:139/99, ks:6/-50, pp:0(0)/0(0), mateNQ: 5, threatN: 15/20, N:
>18982, nmN: 12457, nmNq: 457, qN: 604, cHit: 35%, cqHit: 0%, cUse%: 36, bc%:90,
>pv%:54, pvs%:96,pc%:97,pv:C1C2 G4E3 C2B2 E4E5 B2A2 E3C4
>, D7.0(5): 4.52 : B6D7 , 2.15/99999.0 secs - NPS: 79259 , bf:5.24, avbf:3.3
>AvCut: 0.7 itdeep:717/536, ks:-94/-50, pp:0(0)/0(0), mateN: 9, mateNQ: 88,
>threatN: 111/131, N: 97531, nmN: 63382, nmNq: 5029, qN: 4464, cHit: 63%, cqHit:
>3%, cUse%: 66, bc%:87, pv%:47, pvs%:95,pc%:83,pv:B6D7 G4E3 D2E2 E3F1 C1C2 F1E3
>C2D2
>, D8.0(4): 3.88 : B6D7 , 5.0/99999.0 secs - NPS: 94497 , bf:2.77, avbf:3.17
>AvCut: 0.5 itdeep:1305/930, ks:-94/-50, pp:0(0)/0(0), mateN: 9, mateNQ: 152,
>threatN: 272/403, N: 238190, nmN: 212218, nmNq: 14846, qN: 7229, cHit: 110%,
>cqHit: 3%, cUse%: 116, bc%:87, pv%:55, pvs%:96,pc%:94,pv:B6D7 H4H6 C6C5 B4B5
>C5C4 H6D6 D2D6
>, D9.0(5): 4.35 : B6D7 , 9.67/99999.0 secs - NPS: 98060 , bf:2.01, avbf:2.93
>AvCut: 0.7 itdeep:1619/1099, ks:-94/-50, pp:0(0)/0(0), mateN: 28, mateNQ: 388,
>threatN: 662/1065, N: 458553, nmN: 437865, nmNq: 33063, qN: 18755, cHit: 53%,
>cqHit: 3%, cUse%: 55, bc%:84, pv%:60, pvs%:96,pc%:84,pcc:2,pv:B6D7 F5H6 C1F1
>H4H5 D7F6 H6F7
>, D10.0(6): 3.85 : B6D7 , 22.14/99999.0 secs - NPS: 98781 , bf:2.31, avbf:2.83
>AvCut: 0.7 itdeep:1907/1296, ks:-94/-50, pp:0(0)/0(0), mateN: 29, mateNQ: 591,
>threatN: 2058/3123, N: 897346, nmN: 1174774, nmNq: 81462, qN: 33440, cHit: 84%,
>cqHit: 3%, cUse%: 86, bc%:79, pv%:44, pvs%:95,pc%:94,pcc:21,pv:B6D7 F5H6 F7F6
>H6F7
>, D11.0(6): 4.18 : B6D7 , 69.87/99999.0 secs - NPS: 101529 , bf:3.24, avbf:2.89
>AvCut: 0.8 itdeep:2017/1351, ks:-94/-50, pp:0(0)/0(0), mateN: 29, mateNQ: 3115,
>threatN: 8706/11829, N: 3192200, nmN: 3428197, nmNq: 298184, qN: 175260, cHit:
>55%, cqHit: 2%, cUse%: 57, bc%:80, pv%:52, pvs%:97,pc%:92,pcc:199,pv:B6D7 H4H6
>C6C5 B4C5
>, D12.0(6): 3.9 : B6D7 , 249.58/99999.0 secs - NPS: 101829 , bf:3.58, avbf:2.98
>AvCut: 0.8 itdeep:4407/2572, ks:-94/-50, pp:0(0)/0(0), mateN: 107, mateNQ: 7623,
>threatN: 26643/38472, N: 9955212, nmN: 13709870, nmNq: 1249453, qN: 499897,
>cHit: 60%, cqHit: 2%, cUse%: 61, bc%:77, pv%:54,
>pvs%:94,pc%:-38,pcc:1261,pv:B6D7
>
>
>this is craftys analysis of the position
>(Ra2 was the move my program played in a bullet game - pretty dumb huh!! - hence
>why I use this position to test)
>({10:Mat06} 35. ... Ra2 36. Nf6 Rxg2+ 37. Kxg2 Rc2+ 38. Kg3 Rc3+ 39. Kg4 Rxh3
>40. Rxh3 Nc4 41. Rxh7# +-)
>({10:-5.14} 35. ... Nd7 36. Rh6 Rb1 37. Rd6 Rxd6 38. Nxd6 Kg7 39. Nxb7 h5 40.
>Ne3 Rxb4 -+)
>({10:-1.74} 35. ... Rcc2 36. Nf6 Rxg2+ 37. Kh1 Rh2+ 38. Kg1 Rcg2+ 39. Kf1 Rg7
>40. Nxg7 Kxg7 41. Ne8+ Kg8 42. Nd6 Rb2 43. Rf4 Rb1+ 44. Kg2 Rb2+ 45. Kf3 -+)
>({10:+2.46} 35. ... Nd5 36. exd5 Rcc2 37. Nge3 Rb2 38. Re4 h5 39. dxc6 bxc6 40.
>Ng3 c5 41. bxc5 +-)
>
>
>Thanks
>Scott
>sfarrell@icconsulting.com.au
>sfarrellc@icc
>(you read this far ?)
More like skimmed. I wanted to give you some feedback since I was just working
on IID today. I also use pvs, and I tried many different variations on the
standard IID idea. Recursive and non-recursive, depth-2 vs using a fixed 2 or 3
ply search, some other stuff. I found the occasional position where I'd get a
10% node-reduction, but mostly it was about the same. Not worth bothering with
anymore, for me.
Will
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.