Seer, time-management ... need an update!

[Frank Quisinsky]

Hi there,

Seer is very strong and I like the engine a lot but the time management isn't good.
No time loses but Seer give to many Elo if the moves comes to fast.

Have a look in the results from FCP Qualify Tourney-2021.
OK, move average with 81,6 is not high but I saw to often that Seer have to many time on clock before the new time control started.
Games played with: 40/12, 40 moves in 12 minutes "repeatedly" without resign.

Maybe the programmer have interest to improved that a bit and Seer will win some Elo points.
My personal wish for the next release!

Code: Select all

<pre><b>Engine                 Depth       Time   Games     Moves  Average Forfeit  Book Depth     MIDG   EARLY    ENDG    LATE</b>
Atlas 3.91 POPCNT x6   25.88  134:20:15     320     27397    17.65     0     2994  9.36    21.22 | 22.13 | 24.71 | 37.73
Mr Bob 0.9.0 POPCNT    29.56  122:44:43     320     25488    17.34     0     2699  8.43    24.97 | 25.76 | 30.08 | 37.16
Gogobello 2.2 BMI2 x   28.32  125:39:27     320     26389    17.14     0     3395 10.61    23.31 | 24.00 | 29.18 | 36.87
Stash 24.0 BMI2 x64    30.52  130:06:38     320     27410    17.09     0     2848  8.90    27.87 | 29.05 | 30.59 | 34.93
Amoeba 3.2 x64         32.72  122:31:08     320     26435    16.69     0     2831  8.85    26.75 | 27.91 | 31.56 | 43.89
Hakkapeliitta TCEC v   32.11  123:52:27     320     26756    16.67     0     2813  8.79    26.67 | 27.61 | 32.07 | 42.31
Dirty Cucumber x64     22.67  120:28:08     320     26246    16.52     0     2774  8.67    20.15 | 20.83 | 22.00 | 27.76
FabChess 1.16 BMI2 x   30.36  121:41:20     320     26567    16.49     0     3007  9.40    26.37 | 27.35 | 30.64 | 37.05
Cheng 4.40 dev AVX2    27.92  123:15:09     320     27097    16.37     0     2869  8.97    24.18 | 25.92 | 26.76 | 35.71
Marvin 4.0.1 POPCNT    26.70  119:35:36     320     26397    16.31     0     2982  9.32    23.51 | 24.95 | 24.40 | 33.85
Invictus r305 PEXT x   25.60  115:53:38     320     25717    16.22     0     2832  8.85    22.79 | 23.45 | 25.70 | 31.47
Quazar 0.4 x64         27.54  121:52:13     320     27075    16.20     0     2593  8.10    21.56 | 21.85 | 25.14 | 38.30
Rodent IV 0.30 BMI2    32.15  116:09:28     320     26040    16.06     0     2752  8.60    25.60 | 26.34 | 31.72 | 44.85
Tucano 9.00 x64        31.32  117:35:07     320     26388    16.04     0     2713  8.48    25.51 | 26.96 | 31.08 | 42.05
Nirvanachess 2.4 POP   30.29  112:25:33     320     25542    15.85     0     2872  8.98    26.57 | 27.17 | 30.77 | 38.85
ChessBrainVB 3.74 TC   23.73  120:04:39     320     27440    15.75     0     2838  8.87    21.29 | 22.28 | 22.72 | 30.18
Halogen 8.1 PEXT-AVX   23.41  117:28:59     320     26878    15.74     0     2844  8.89    21.91 | 22.72 | 24.64 | 24.45
Lc0 0.26.3 x64          7.43  120:39:43     320     27730    15.66     0     2933  9.17     8.89 |  9.68 |  7.04 |  5.50
Topple 0.7.5 Skylake   30.34  118:26:30     320     27277    15.63     0     3017  9.43    27.42 | 28.16 | 29.23 | 36.40
Orion 0.8 NN POP AVX   22.59  116:06:02     320     26893    15.54     0     2945  9.20    19.62 | 20.50 | 22.73 | 27.99
Koivisto 4.0 POPCNT    33.99  115:11:51     320     26700    15.53     0     5858 18.31    28.74 | 29.43 | 33.31 | 46.62
SmarThink 1.98 AVX2    24.02  118:49:14     320     27772    15.40     0     2733  8.54    22.75 | 24.20 | 23.78 | 26.48
chess22k 1.14 JAVA x   34.08  111:08:55     320     26135    15.31     0     2790  8.72    26.85 | 27.12 | 34.97 | 48.62
Bagatur 2.2 JAVA x64   40.51  117:42:30     320     27827    15.23     0     2786  8.71    29.13 | 29.90 | 40.73 | 59.46
Cheese 2.2 POPCNT x6   22.11  112:55:51     320     26753    15.20     0     2805  8.77    20.22 | 21.05 | 21.30 | 26.64
Gödel 7.0 SSE42 x64    27.84  113:51:07     320     27277    15.03     0     3389 10.59    23.99 | 24.64 | 26.22 | 38.62
Gaviota 1.0 AVX x64    19.27  108:04:51     320     26050    14.94     0     3131  9.78    18.38 | 19.33 | 20.08 | 19.81
Monolith 2.01 PEXT x   28.49  106:52:16     320     25982    14.81     0     2878  8.99    24.67 | 25.42 | 27.91 | 36.69
pirarucu 3.3.5 JAVA    32.25  111:20:52     320     27180    14.75     0     2868  8.96    27.52 | 27.97 | 32.75 | 41.03
TheBaron 3.44.1 x64    18.92  106:26:56     320     26158    14.65     0     3059  9.56    16.78 | 17.70 | 18.04 | 23.87
Wasp 4.08 dev Modern   29.37   97:40:44     320     24209    14.53     0     2908  9.09    25.08 | 26.30 | 26.58 | 42.41
Weiss 1.2 PEXT x64     27.37  105:12:21     320     26388    14.35     0     2767  8.65    27.00 | 27.66 | 28.44 | 26.70
Hiarcs 14 WCSC w32     23.33  107:12:41     320     26979    14.31     0     2822  8.82    19.73 | 20.84 | 23.94 | 28.42
Naum 4.6 x64           20.43  107:30:35     320     27157    14.25     0     2858  8.93    18.15 | 19.17 | 20.76 | 25.05
Junior 13.3.00 x64     25.82  108:00:12     320     27490    14.14     0     2859  8.93    18.63 | 19.36 | 24.14 | 41.11
Protector 1.9.0 x64    27.96  100:11:47     320     25571    14.11     0     2817  8.80    20.70 | 21.96 | 23.65 | 45.08
Spike 1.4 Leiden w32   21.23  102:41:28     320     26402    14.00     0     2807  8.77    20.00 | 20.97 | 21.62 | 23.11
Crafty 25.6 x64        26.27   96:53:42     320     25656    13.60     0     2884  9.01    26.08 | 26.79 | 27.98 | 23.97
Asymptote 0.8 Broadw   22.92   97:32:38     320     26091    13.46     0     2888  9.02    21.88 | 23.01 | 22.61 | 25.19
Francesca 0.29a x64    33.78   96:00:45     320     25686    13.46     0     2944  9.20    31.57 | 32.77 | 33.91 | 38.14
Seer 1.2.1 NNUE Skyl   23.81   80:24:11     320     25200    11.49     0     2786  8.71    22.82 | 23.48 | 24.64 | 24.59


Legend

<b>Depth</b>     : Overall average search depth
<b>Time</b>      : Total time engine used
<b>Moves</b>     : Total moves engine played
<b>Average</b>   : Average time per move in centi-seconds
<b>Forfeit</b>   : Games engine lost due to time forfeit
<b>Book</b>      : Moves played from opening book plus average depth

<b>MIDG</b>      : Average search depth during the opening and middle game
<b>EARLY</b>     : Average search depth during the middle game without Queens, the so called early end game
<b>ENDG</b>      : Average search depth during the end game
<b>LATE</b>      : Average search depth during the late end game

List is sorted on <i>Average Time</i> indicating the engine that uses the most time tops.

Best
Frank

[connor_mcmonigle]

Thanks for the information and including Seer in your tournaments! I've been procrastinating on improving the time management for some time now (haha, perhaps I need to improve my own time management). Time management is definitely something I'll explore before the next release.

The current dev version is already substantially stronger (~+110 elo in my testing) than v1.2.1 through a number of improvements to the search. I'm currently working on training new networks "from zero" and observing some promising results. If all goes according to plan, I should have a new release ready in the next few weeks and will definitely revisit the time management before then.

[Frank Quisinsky]

need to improve your own time management ...


My time management is very bad.
Better times will be come ...


Hi Conner,

if you like ...
Send me your current engine version and I will replay the games from the still-running FCP Tourney-2021.
During round 1 all is possible, but if we created the first stats with Excel no changes are possible (to many work).

Again, if you like!
Can be send with my contact form (impressum):

With the results:
Frank is happy, you have more interesting results and the follower from the tourney will be happy too.

Frank's Chesss Page:
https://www.amateurschach.de

Best
Frank

[Frank Quisinsky]

http://www.amateurschach.de/main/_contact.htm
The direct link!

Best
Frank

[connor_mcmonigle]

Thanks for the offer, but I'd rather just continue with v1.2.1 participating in your tournament

Realistically, my new "from zero" networks will be substantially weaker than the network bundled with v1.2.1. However, I'm willing to accept that regression as I want Seer to exist completely independent from other engines (current networks are trained with a low depth Stockfish search as a target). The many search improvements I've made give me a roughly 100 elo cushion to work with while still insuring the next release is considerably stronger than v1.2.1. I'd rather not have stronger, unofficial, versions floating around and, therefore, would prefer to not submit a compile of the latest dev version.

[Frank Quisinsky]

Hi Conner,

so, let me test it if ready with the next FCP Tourney, start middle of the year.



Fact is, your engine is qualified and in the group of TOP-41 and the style is great.
Produced nice games and the move average is very low (I think a little secret in your development to hold such a low move average). With FCP Tourney-2021 and stronger opponents the situation can be a complete other as with FCP Qualify Tourney-2021.

So, good luck for your engine and thanks for the explanation.

Go Seer go ...



Best
Frank