OpenChess

I’m new to this, so don’t be too hard on me for asking a stupid question that may have been answered previously (I sure can’t find it). Robert Houdart is probably the only one who can truly answer this, but…

Why does changing the hash table size in Houdini 1.5a also change the recommended move in deep analysis at equivalent base depths?

Perhaps I’m interpreting the output wrong?

32M hash table:
FEN: r2qkb1r/pb1n1ppp/2p1pn2/1p6/3P4/2NBPN2/PP3PPP/R1BQK2R w KQkq - 2 9

Houdini_15a_w32:
4/12 00:00 1,287 0 +0.61 O-O b5-b4 Nc3-e4 Nf6xe4 Bd3xe4
5/13 00:00 2,052 0 +0.47 O-O b5-b4 Nc3-e4 Nf6xe4 Bd3xe4 Nd7-f6
5/20 00:00 3,185 0 +0.59 a2-a3 b5-b4 a3xb4 Bf8xb4 O-O O-O
6/29 00:00 7,733 220,000 +0.55 a2-a3 b5-b4 a3xb4 Bf8xb4 O-O O-O Qd1-b3
7/29 00:00 10,781 256,000 +0.54 a2-a3 b5-b4 a3xb4 Bf8xb4 O-O O-O h2-h3 Qd8-c7 Bc1-d2 c6-c5
8/29 00:00 20,966 343,000 +0.46 a2-a3 b5-b4 a3xb4 Bf8xb4 O-O O-O h2-h3 c6-c5 Bc1-d2 c5xd4 e3xd4
9/29 00:00 50,891 431,000 +0.40 a2-a3 b5-b4 a3xb4 Bf8xb4 O-O O-O Qd1-b3 a7-a5 Bc1-d2 c6-c5 Rf1-c1 Qd8-c7
10/29- 00:00 81,568 463,000 +0.33 a2-a3 b5-b4
10/29 00:00 158,504 496,000 +0.37 a2-a3 b5-b4 a3xb4 Bf8xb4 O-O O-O Qd1-b3 a7-a5 Bc1-d2 c6-c5 Rf1-c1 Bb7xf3 g2xf3 Nd7-e5
11/29 00:00 216,108 510,000 +0.40 a2-a3 c6-c5 d4xc5 Bf8xc5 b2-b4 Bc5-e7 Nc3xb5 a7-a6 Nb5-d4 O-O O-O Nf6-e4
12/30- 00:00 473,871 537,000 +0.33 a2-a3 c6-c5
12/30 00:01 700,683 539,000 +0.36 a2-a3 b5-b4 a3xb4 Bf8xb4 O-O c6-c5 Bd3-a6 Bb7xa6 Ra1xa6 O-O h2-h3 c5xd4 e3xd4 Qd8-c7 Qd1-d3 h7-h6
13/30 00:01 824,885 541,000 +0.33 a2-a3 b5-b4 a3xb4 Bf8xb4 O-O c6-c5 Bd3-a6 Bb7xa6 Ra1xa6 O-O Nf3-e5 c5xd4 e3xd4 Nd7xe5 d4xe5 Nf6-d5 Nc3xd5 Qd8xd5 Qd1xd5 e6xd5
14/31 00:02 1,410,837 548,000 +0.29 a2-a3 b5-b4 a3xb4 Bf8xb4 O-O c6-c5 Bc1-d2 O-O Qd1-e2 c5xd4 e3xd4 Nd7-b6 Bd2-e3 Nb6-d5 Nc3xd5 Qd8xd5 Be3-f4 Qd5-d7
15/31+ 00:03 2,117,881 549,000 +0.35 a2-a3
15/41 00:06 3,672,794 550,000 +0.25 a2-a3 b5-b4 a3xb4 Bf8xb4 Bc1-d2 c6-c5 O-O O-O Qd1-e2 c5xd4 e3xd4 Nd7-b6 Bd2-e3 Nb6-d5 Rf1-c1 Nd5xe3 f2xe3 a7-a5 Nf3-e5 Bb4-d6 Nc3-b5 Bd6xe5 d4xe5
16/41 00:10 5,933,082 549,000 +0.25 a2-a3 b5-b4 a3xb4 Bf8xb4 Bc1-d2 c6-c5 O-O O-O Qd1-e2 c5xd4 e3xd4 Nd7-b6 Bd2-e3 Nb6-d5 Rf1-c1 Nd5xe3 f2xe3 a7-a5 Nf3-e5 Bb4-d6 Nc3-b5 Bd6xe5 d4xe5
17/45 00:23 12,883,663 555,000 +0.21 a2-a3 b5-b4 a3xb4 Bf8xb4 Bc1-d2 c6-c5 O-O O-O Qd1-e2 a7-a5 Rf1-c1 Qd8-b6 Nc3-a4 Qb6-d6 Bd2xb4 Bb7xf3 g2xf3 c5xb4 Na4-c5 Rf8-c8 Nc5-b7 Rc8xc1+ Ra1xc1
18/45 00:51 29,268,899 566,000 +0.20 a2-a3 b5-b4 a3xb4 Bf8xb4 Bc1-d2 c6-c5 O-O O-O Qd1-e2 a7-a5 Rf1-d1 Ra8-c8 Nc3-b5 Bb4xd2 Nf3xd2 Qd8-b6 Rd1-c1 c5xd4 e3xd4 a5-a4 h2-h3 h7-h6 Rc1xc8 Rf8xc8 Ra1xa4 Rc8-c1+ Kg1-h2
19/45 01:33 53,913,489 575,000 +0.20 a2-a3 b5-b4 a3xb4 Bf8xb4 O-O c6-c5 Bd3-a6 Bb7xa6 Ra1xa6 O-O Qd1-e2 c5xd4 e3xd4 Nd7-b8 Ra6-a1 Nb8-c6 Bc1-e3 a7-a5 Rf1-c1 Ra8-c8 h2-h3 Qd8-d7 Nf3-e5 Nc6xe5 d4xe5 Bb4xc3 b2xc3 Nf6-d5 Ra1xa5 Nd5xc3
20/45 02:19 80,770,939 580,000 +0.23 a2-a3 b5-b4 a3xb4 Bf8xb4 O-O c6-c5 Bd3-a6 Bb7xa6 Ra1xa6 O-O Qd1-e2 c5xd4 e3xd4 Nd7-b8 Ra6-a1 Nb8-c6 Bc1-e3 a7-a5 Rf1-c1 Ra8-c8 h2-h3 Qd8-d6 Nc3-b5 Qd6-d5 Be3-f4 h7-h6
21/45 04:22 155,169,813 591,000 +0.20 a2-a3 b5-b4 a3xb4 Bf8xb4 O-O c6-c5 Qd1-e2 O-O Bc1-d2 a7-a5 Ra1-c1 Ra8-c8 Rc1-c2 Qd8-b6 Rf1-c1 c5xd4 e3xd4 Bb7xf3 Qe2xf3 Qb6xd4 Nc3-b5 Rc8xc2 Nb5xd4 Rc2xd2 Nd4-c6 Rd2xb2 Qf3-g3 Rf8-c8
22/54 15:15 558,192,851 609,000 +0.18 a2-a3 b5-b4 a3xb4 Bf8xb4 O-O c6-c5 Bd3-a6 Bb7xa6 Ra1xa6 O-O Qd1-e2 c5xd4 e3xd4 Nd7-b8 Ra6-a4 a7-a5 Bc1-g5 Nb8-d7 Ra4-a1 Ra8-c8 Rf1-c1 Qd8-b6 Bg5-f4 Bb4xc3 b2xc3
23/54 23:39 869,792,366 612,000 +0.18 a2-a3 b5-b4 a3xb4 Bf8xb4 O-O c6-c5 Bd3-a6 Bb7xa6 Ra1xa6 O-O Qd1-e2 c5xd4 e3xd4 Nd7-b8 Ra6-a4 a7-a5 Bc1-g5 Nb8-d7 Rf1-c1 Ra8-c8 Ra4-a1 Qd8-b6 Bg5-f4 Bb4xc3 b2xc3 Nf6-d5 Qe2-d2 Nd7-f6
23/60 34:25 1,268,888,686 614,000 +0.19 e3-e4 b5-b4 Nc3-a4 c6-c5 e4-e5 Nf6-d5 Na4xc5 Bf8xc5 d4xc5 Qd8-c7 O-O Nd7xc5 Bd3-b5+ Bb7-c6 Nf3-d4 Bc6-d7 Qd1-e2 O-O Bc1-g5 h7-h6 Bg5-d2 Ra8-c8 Ra1-c1 Bd7xb5 Nd4xb5 Qc7-a5 Nb5-d6 Qa5xa2 Nd6xc8 Rf8xc8 Rc1-c2
24/60 48:44 1,807,720,685 618,000 +0.20 e3-e4 b5-b4 Nc3-a4 c6-c5 e4-e5 Nf6-d5 Na4xc5 Bf8xc5 d4xc5 Qd8-a5 Qd1-e2 Nd7xc5 Bd3-b5+ Ke8-f8 O-O Nc5-b3 a2xb3 Qa5xa1 Nf3-g5 h7-h6 Ng5-e4 Ra8-c8 Bb5-c4 Kf8-g8 Ne4-d6 Rc8-c7 Rf1-d1
25/65 1:09:39 2,604,808,396 623,000 +0.19 e3-e4 b5-b4 Nc3-a4 c6-c5 e4-e5 Nf6-d5 Na4xc5 Bf8xc5 d4xc5 Qd8-c7 O-O Nd7xc5 Bd3-b5+ Bb7-c6 Nf3-d4 Bc6-d7 Qd1-e2 O-O Bc1-d2 Qc7-b6 Ra1-c1 Ra8-c8 Bb5-c4 Nc5-a4 Bc4xd5 Rc8xc1 Rf1xc1 Qb6xd4 Bd5-e4
25/71 1:38:38 3,729,920,567 630,000 +0.19 e3-e4 b5-b4 Nc3-a4 c6-c5 e4-e5 Nf6-d5 Na4xc5 Bf8xc5 d4xc5 Qd8-c7 O-O Nd7xc5 Bd3-b5+ Bb7-c6 Nf3-d4 Bc6-d7 Qd1-e2 O-O Bc1-g5 Qc7-b6 Bb5-c4 h7-h6 Bg5-d2 Nc5-a4 Nd4-f5 e6xf5
Best move: e3-e4, Value: +0.19, Depth: 25/71, Time: 1:38:38.104, 3,729,920,567 Nodes, 630 kN/sec.

1024M hash table:
FEN: r2qkb1r/pb1n1ppp/2p1pn2/1p6/3P4/2NBPN2/PP3PPP/R1BQK2R w KQkq - 2 9

Houdini_15a_w32:
4/12 00:00 1,287 0 +0.61 O-O b5-b4 Nc3-e4 Nf6xe4 Bd3xe4
5/13 00:00 2,052 0 +0.47 O-O b5-b4 Nc3-e4 Nf6xe4 Bd3xe4 Nd7-f6
5/20 00:00 3,185 0 +0.59 a2-a3 b5-b4 a3xb4 Bf8xb4 O-O O-O
6/29 00:00 7,733 193,000 +0.55 a2-a3 b5-b4 a3xb4 Bf8xb4 O-O O-O Qd1-b3
7/29 00:00 10,781 229,000 +0.54 a2-a3 b5-b4 a3xb4 Bf8xb4 O-O O-O h2-h3 Qd8-c7 Bc1-d2 c6-c5
8/29 00:00 20,966 312,000 +0.46 a2-a3 b5-b4 a3xb4 Bf8xb4 O-O O-O h2-h3 c6-c5 Bc1-d2 c5xd4 e3xd4
9/29 00:00 50,891 400,000 +0.40 a2-a3 b5-b4 a3xb4 Bf8xb4 O-O O-O Qd1-b3 a7-a5 Bc1-d2 c6-c5 Rf1-c1 Qd8-c7
10/29- 00:00 81,568 438,000 +0.33 a2-a3 b5-b4
10/29 00:00 158,504 471,000 +0.37 a2-a3 b5-b4 a3xb4 Bf8xb4 O-O O-O Qd1-b3 a7-a5 Bc1-d2 c6-c5 Rf1-c1 Bb7xf3 g2xf3 Nd7-e5
11/29 00:01 216,108 486,000 +0.40 a2-a3 c6-c5 d4xc5 Bf8xc5 b2-b4 Bc5-e7 Nc3xb5 a7-a6 Nb5-d4 O-O O-O Nf6-e4
12/30- 00:01 473,863 517,000 +0.33 a2-a3 c6-c5
12/30 00:01 700,645 520,000 +0.36 a2-a3 b5-b4 a3xb4 Bf8xb4 O-O c6-c5 Bd3-a6 Bb7xa6 Ra1xa6 O-O h2-h3 c5xd4 e3xd4 Qd8-c7 Qd1-d3 h7-h6
13/30 00:02 830,174 523,000 +0.33 a2-a3 b5-b4 a3xb4 Bf8xb4 O-O c6-c5 Bd3-a6 Bb7xa6 Ra1xa6 O-O Nf3-e5 c5xd4 e3xd4 Nd7xe5 d4xe5 Nf6-d5 Nc3xd5 Qd8xd5 Qd1xd5 e6xd5
14/34- 00:03 1,659,571 536,000 +0.27 a2-a3 b5-b4
14/37 00:06 3,212,387 540,000 +0.25 a2-a3 b5-b4 a3xb4 Bf8xb4 O-O c6-c5 Qd1-e2 O-O Rf1-d1 a7-a5 Bc1-d2 Qd8-b6 Ra1-c1 Ra8-c8 Nc3-b5 Bb4xd2 Qe2xd2 Bb7xf3 g2xf3 c5xd4 Nb5xd4
15/37 00:09 4,898,515 541,000 +0.21 a2-a3 b5-b4 a3xb4 Bf8xb4 O-O c6-c5 Qd1-e2 O-O h2-h3 a7-a5 Bc1-d2 Qd8-b6 Rf1-d1 c5xd4 e3xd4 Ra8-c8 Bd2-f4 Bb4xc3 b2xc3 Rc8xc3
16/37 00:14 7,833,631 542,000 +0.26 a2-a3 b5-b4 a3xb4 Bf8xb4 O-O c6-c5 Qd1-e2 O-O Bc1-d2 Qd8-b6 Rf1-c1 Bb7xf3 g2xf3 a7-a6 Nc3-a4 Qb6-d6 Bd2xb4 c5xb4 f3-f4 Rf8-c8 Rc1xc8+ Ra8xc8 Bd3xa6
17/37 00:20 10,915,460 545,000 +0.26 a2-a3 b5-b4 a3xb4 Bf8xb4 O-O c6-c5 Qd1-e2 O-O Bc1-d2 Qd8-b6 Rf1-c1 Bb7xf3 g2xf3 a7-a6 Nc3-a4 Qb6-d6 Bd2xb4 c5xb4 f3-f4 Rf8-c8 Rc1xc8+ Ra8xc8 Bd3xa6
18/37 00:29 16,216,181 552,000 +0.26 a2-a3 b5-b4 a3xb4 Bf8xb4 O-O c6-c5 Qd1-e2 O-O Bc1-d2 Qd8-b6 Rf1-c1 Bb7xf3 g2xf3 a7-a6 Nc3-a4 Qb6-d6 Bd2xb4 c5xb4 f3-f4 Rf8-c8 Rc1xc8+ Ra8xc8 Bd3xa6
19/47 00:57 31,925,735 556,000 +0.25 a2-a3 b5-b4 a3xb4 Bf8xb4 O-O c6-c5 Qd1-e2 O-O Bc1-d2 Qd8-b6 Rf1-d1 Ra8-c8 Bd3-a6 c5xd4 e3xd4 Rf8-d8 Rd1-c1 Bb4-d6 Ba6xb7 Qb6xb7 Nc3-b5 Bd6-b8 g2-g3 Nd7-b6
20/47 01:49 61,106,519 560,000 +0.25 a2-a3 b5-b4 a3xb4 Bf8xb4 O-O c6-c5 Qd1-e2 O-O Bc1-d2 Qd8-b6 Rf1-d1 Ra8-c8 Bd3-a6 c5xd4 e3xd4 Rf8-d8 Rd1-c1 Bb4-d6 Ba6xb7 Qb6xb7 Nc3-b5 Bd6-b8 g2-g3 Nd7-b6
21/53 03:59 135,328,197 566,000 +0.24 a2-a3 b5-b4 a3xb4 Bf8xb4 O-O c6-c5 Qd1-e2 O-O Bc1-d2 Qd8-b6 Nc3-a4 Qb6-d6 d4xc5 Bb4xc5 Rf1-d1 Bb7xf3 g2xf3 Bc5-b4 Bd2-c3 Nf6-d5 Bd3-e4 Ra8-b8 Qe2-d3 Bb4xc3 Be4xh7+ Kg8-h8 Na4xc3 Rb8xb2
22/54- 08:22 288,208,882 574,000 +0.17 a2-a3 b5-b4
22/54 14:38 504,496,529 574,000 +0.18 a2-a3 b5-b4 a3xb4 Bf8xb4 O-O c6-c5 Qd1-e2 O-O Bc1-d2 a7-a5 Rf1-c1 Qd8-b6 Nc3-a4 Qb6-d6 d4xc5 Bb4xc5 Bd2-c3 Bc5-b4 Rc1-d1 Bb7xf3 Qe2xf3 Nd7-e5 Bc3xe5 Qd6xe5 Ra1-c1 Ra8-b8 Qf3-f4 Qe5xf4 e3xf4
23/64 27:22 945,443,429 575,000 +0.20 a2-a3 Bf8-d6 O-O a7-a5 Qd1-c2 O-O Nf3-g5 Bd6xh2+ Kg1xh2 Nf6-g4+ Kh2-g1 Qd8xg5 f2-f3 Ng4-f6 Bc1-d2 e6-e5 Nc3-e4 Nf6xe4 Bd3xe4 Nd7-f6 d4xe5 Qg5xe5 Bd2-c3 Qe5-c5 Bc3-d4 Qc5xc2 Be4xc2 b5-b4 Rf1-d1 Rf8-d8
24/64 54:11 1,890,498,531 581,000 +0.16 a2-a3 Bf8-d6 O-O O-O Qd1-c2 a7-a5 Nf3-g5 h7-h6 Ng5-e4 Nf6xe4 Nc3xe4 Bd6-e7 Bc1-d2 c6-c5 Ne4xc5 Nd7xc5 d4xc5 Be7xc5 Ra1-d1 Qd8-d5 e3-e4 Qd5-c6 Bd2-c3 Rf8-c8 Bc3-e5 Qc6-d7
25/64 1:33:48 3,303,319,991 586,000 +0.20 a2-a3 Bf8-d6 O-O O-O Qd1-c2 a7-a5 Nf3-g5 h7-h6 Ng5-e4 Nf6xe4 Nc3xe4 Bd6-e7 Rf1-d1 Qd8-b6 Bc1-d2 b5-b4 Rd1-c1 b4-b3 Qc2-c4 a5-a4 Bd3-b1 Qb6-c7 f2-f4
25/64 2:15:47 4,834,784,916 593,000 +0.20 a2-a3 Bf8-d6 O-O O-O Qd1-c2 a7-a5 Nf3-g5 h7-h6 Ng5-e4 Nf6xe4 Nc3xe4 Bd6-e7 Rf1-d1 Qd8-b6 Bc1-d2 b5-b4 Rd1-c1 b4-b3 Qc2-c4 a5-a4 Bd3-b1 Qb6-c7 f2-f4
Best move: a2-a3, Value: +0.20, Depth: 25/64, Time: 2:15:47.464, 4,834,784,916 Nodes, 593 kN/sec.

Please, help me understand.

Is this single CPU? Otherwise the Multi-CPU indeterminism is going to change the move.

One poor little old Intel P4, single thread setting for Houdini.

MoldyJacket wrote: single thread setting for Houdini.

All this is perfectly normal, with any program:
analysis is not the same for different hashtable size.
Here you have 2 moves with practically the same evaluation.

Try the same infinite analysis, this time with 2 PVs (Multi-PV mode), you will see how close they are...

Thank you, it is good to know that many engines change their evaluation based on hash size. This seems to be an important consideration that goes well beyond simply tuning for particular hardware.

Granted +0.16 and +0.20 is practically the same in human terms, but I assumed it would be very significant to a chess engine. Base depth 25 can’t be considered since neither run fully completed it.

I started out one pleasant evening with the thought I’d like to start playing chess again, so I jumped on the Internet to see what tools may be available these days. Wow, I found a Renaissance in the setting of the Wild West!

I settled on an engine and prepared to get my ass kicked repeatedly in an effort to improve my game and explore ideas. I reasoned the only thing I could truly hang my hat on was the single selected move of completed and equivalent base depths to tune my old machine.

Now I have wasted a fair amount of kilowatts exploring different engines and how they change their evaluation based purely on different hash sizes. Absolutely fascinating, but I’m not necessarily improving my game!

I suppose an engine can use the space anyway it likes, but I assumed it was merely cache to speed the search and not part of the evaluation mechanism itself. I understand how the scoring of possible moves can be relative to a particular search path, but the final result should be tightly bound…shouldn’t it?

In the end I’ve draw the conclusion you must be satisfied to play a different game depending on your choice of hash size, even if more time is given to counterbalance platform constraints. And who can truly say more is better without imperial evidence or special knowledge anyway?

Well, I can’t edit my own post. Obviously I meant say “drawn” (not “draw”) and empirical (not “imperial”).

Just in case anyone is interested, I did get the definitive answer to my question courtesy of Dr. Hyatt: Sub-Tree Grafting. Thank you!

http://talkchess.com/forum/viewtopic.ph ... 34&t=38597

I tested Houdini, seems it will only use 1024 MB (32-bit) maximum as documented.

MoldyJacket wrote:I did get the definitive answer to my question courtesy of Dr. Hyatt: Sub-Tree Grafting.

Well, good for you!
Now, did you even understand that answer? I sure didn't...

Sure, more is virtually always better for two reasons: Search efficiency (speed obviously) and accuracy (deeper unintended depths from grafting).

Basically, it is a waste not filling the hash before a result is required (short time controls). But, it is potentially an even bigger waste filling the hash too soon and getting more overwrites. If you care about analysis and long time controls, you want all the hash your system and code can stand.

I haven’t found a clearer and concise explanation of the concept in layman’s terms on the Web yet, including that Chess Programming Wiki.