On-line engine blitz tourney August

[Joost Buijs]

The monthly on-line engine blitz tourney for August will take place on:

Saturday August 14, 2:00 PM EST (Boston time), 20:00 CET (Amsterdam time)

To connect:

winboard -zp -ics -icshost nightmare-chess.nl -icshelper timeseal -fcp ENGINE.exe -fd ENGINEFOLDER -autoKibitz -keepAlive 30

(for UCI engines, add -fUCI)

People without account can submit a request for an account to admin@nightmare-chess.nl

Our tournament manager will be: Eendje (Lisebeth at ICC)

The format will be 9 rounds Swiss 5/1

Lisebeth likes to start the tournament on top of the hour, please try to be on-line 15 minutes prior to tournament start, late joins are not supported.

[amanjpro]

A multithreaded Zahak will participate hopefully... Two threads for sure, probably 4 (or even 8?) my tests need to finish first to know

[Joost Buijs]

This time there were less participants than usual, maybe some people are on vacation. HGM's engines are still missing too.
Myrddin logged on almost 30 minutes too late. Skiull didn't move at all and lost all games.

Nightmare won the tournament like last month, again it was a bit lucky, this time with black against Arasan. Arasan used a brand new implementation of NNUE evaluation, probably this still has to be optimized.

Nemorino came in second and Wasp ended third. Wasp used an experimental version of a rather small NNUE network, it seems to do better than John expected.

Zahak improved a lot, it looks like it did 150 Elo better than last month.

Code: Select all

Tourney Players: Round 9 of 9 
:
:     Name              Rating Score Perfrm Upset  Results 
:     ----------------- ------ ----- ------ ------ ------- 
:  1 +NightmareX        [2913]  7.5  [2927] [   0] +10w =03w +11b =06w +07b =02w +04b +05w +09b 
:  2 +Nemorino          [2885]  7.0  [2913] [  14] +13w +04w =03b =05w +09b =01b +07w =06b +10b 
:  3 +WaspX             [2811]  6.5  [2839] [ 110] +15w =01b =02w +12b =06b =05w +08w =04b +07w 
:  4 +Marvin            [2737]  6.0  [2757] [ 178] +16w -02b =05w +08b +12w +06b -01w =03w +14b 
:  5 +ArasanX           [2839]  5.5  [2700] [  23] +12w =08w =04b =02b +15w =03b =06w -01b +16w 
:  6 +Goldbar           [2827]  5.5  [2664] [  78] +14w +09w =08b =01b =03w -04w =05b =02w +17b 
:  7 +WaDuuttie         [2610]  5.0  [2563] [   0] =11w =10b +14w +15b -01w +09w -02b +13w -03b 
:  8 +atomNightmare     [2585]  5.0  [2529] [ 248] +18w =05b =06w -04w =11b =10w -03b +16b +13w
:  9 +atomGoldbar       [2576]  5.0  [2523] [   0] +17w -06b +13b +10w -02w -07b +14b +12w -01w 
: 10 +Arminius          [2552]  5.0  [2428] [  45] -01b =07w +17b -09b +14w =08b +18w +11b -02w 
: 11 +Zahak             [2106]  5.0  [2462] [1045] =07b +13w -01w =14b =08w =12b +15b -10w +18w 
: 12 -RookieMonster     [2330]  4.5  [2337] [   0] -05b +18w +16b -03w -04b =11w +17b -09b +15w 
: 13 +Nemeton           [2356]  4.0  [2258] [   0] -02b -11b -09w +18w +17b +15w +16w -07b -08b 
: 14 +rpiArminius       [2238]  3.5  [2260] [   0] -06b +17w -07b =11w -10b +16b -09w +18b -04w 
: 15 +Barbarossa        [2232]  3.0  [2176] [   0] -03b +16w +18b -07w -05b -13b -11w +17w -12b 
: 16 +CTdccc            [2188]  2.0  [2095] [   0] -04b -15b -12w +17w +18b -14w -13b -08w -05b
: 17 -Skipper           [1597]  1.0  [2049] [ 350] -09b -14b -10w -16b -13w +18b -12w -15b -06w 
: 18 -Skiull            [1947]  0.0  [1842] [   0] -08b -12b -15w -13b -16w -17w -10b -14w -11b 
:
:     Average Rating    2462.7

[jdart]

Nightmare won the tournament like last month, again it was a bit lucky, this time with black against Arasan.

Arasan was undefeated, losing only to Nightmare. It had a bunch of draws. So not bad. In the Nightmare game, I think it just choose a poor opening line (1. d4 Nf6 2. c4 e6 3. g3 d5 4. Bg2 dxc4 5. Nf3 Bd7?!), rather than some major error.

[jstanback]

I was very surprised by the results for Wasp NN as it is about 100 Elo worse than Wasp 4.5 at Game/7s + 0.07s. The NN gives over 2X slowdown in NPS as compared to HCE, so I expected it would not be quite so bad when given more than 1s per move. Maybe at 26 threads it's actually getting close to equal with Wasp 4.5? Too few games for me to tell anything other than it's doing well enough to keep me interested in improving it.

This NN has 2x12x64 PSQ inputs (one set for both kings on same flank, another set when on different flanks), 2 inputs for side to move, and 38 inputs for material difference. It has two hidden layers, the first has 96 nodes and the second has 32 nodes. For training, I use the same set of 52M positions taken from Wasp games as I use for tuning my HCE. But for the NN about 75% of the training is done by selecting a random position from this set and than doing a random playout from it to get a lot of "junk" positions. For these, the HCE is used as the target. The other 25% of the training uses positions randomly chosen from the 52M pos set, with target set mostly from the game result. I trained this NN with a total of 25B positions. I still have a lot of experimenting to do to try to improve my training methodology.

John

[amanjpro]

Zahak improved a lot, it looks like it did 150 Elo better than last month.

Thanks a lot for the tourney... While Zahak has improved a bit (some 40 elo), the rest comes from additional threads (thanks to the newly implemented LazySMP)

But, of course I am super happy with the result!

[Joost Buijs]

Nightmare won the tournament like last month, again it was a bit lucky, this time with black against Arasan.

Arasan was undefeated, losing only to Nightmare. It had a bunch of draws. So not bad. In the Nightmare game, I think it just choose a poor opening line (1. d4 Nf6 2. c4 e6 3. g3 d5 4. Bg2 dxc4 5. Nf3 Bd7?!), rather than some major error.

To me the position seemed rather equal and I was surprised to see the score slowly rising up to the point where Arasan lost a pawn on the queen wing.

I'm also surprised to see how good the NN of Nightmare does with 32 threads on a fast computer. It's search is pretty basic without all the fancy stuff they use nowadays.

For move sorting it basically has: TT, 2 killers, 1 counter-move and history.
For pruning it uses: razoring, futility, null-move and LMR.
Extended are check-evasions and it has singular-extensions.

The network uses 2x 768 inputs and has 384x32x1 neurons. It could do with 2x 736 inputs but I wanted to have some extra inputs for castling and other stuff like that. The whole inference code is 16 bit, I've tried 8 bit in the past, it was worse and not faster at all. I'm also working on an international draughts NNUE engine, it gives me exactly the same results (8 bit not being much faster than 16).

Training networks is very time consuming, usually networks with the lowest MSE are not the best, it's really worthwhile to test a network after each epoch, this is what takes so much time.