Tried testing it with UCI_Elo 1500 at TC 60s+100ms and it losses on time forfeit.Frank Quisinsky wrote: ↑Fri Sep 06, 2019 6:09 pm Hi there,
Please be informed that John Stanback released Wasp 3.75!
Frank's Chess Page:
http://www.amateurschach.de
All the important changes can be found in release-notes.txt, included in 2.694Mb *.zip file.
Can be find on Wasp news selection too: http://www.amateurschach.de/main/_wasp.htm
Please read the news before installation!
Polyglot book format, no longer Wasp.rc and some other things are changed!
Since September 01st, 2009 a test-run vs. 20 opp. is still running (7x stronger, 6x equal, 7x weaker).
So far 904 / 1.000 games played. More 1.000er test-runs will be follow.
My test-runs will be run not longer in "LIVE" mode.
From time to time I updated to the currrent results and games.
Have fun with Wasp 3.75 by John!
And ...
Thanks John!
Best
Frank
Looks more like a hack than a solution.jstanback wrote: ↑Sat Sep 07, 2019 12:25 am I made a change where if the first move at the root fails low several times (score has dropped about 2 pawns or more) then I accept that score for the first move, set a flag, and search the remaining moves to see if one of them is better. If not, then I repeat the iteration.
Hi Ferdy,
I saw no Elo gain or loss after this change. But I think it's a good solution to the problem of the first move at the root failing low by a big margin. Wasp sometimes wastes a lot of time to resolve just how bad the first move is, but with this change it will try the other moves to see if any of them are better. If so, things are good and it proceeds to the next iteration (or time out). Otherwise, it will redo the current iteration starting with the bad first move and a lowered alpha-beta window. In test positions and looking at logs from games I think this is working well.lucasart wrote: ↑Sat Sep 14, 2019 2:39 amLooks more like a hack than a solution.jstanback wrote: ↑Sat Sep 07, 2019 12:25 am I made a change where if the first move at the root fails low several times (score has dropped about 2 pawns or more) then I accept that score for the first move, set a flag, and search the remaining moves to see if one of them is better. If not, then I repeat the iteration.
Did this prove to gain elo in your testing ?
Perhaps it makes sense to widen asymmetrically, widening faster on fail lows than fail highs.
Tried the ucielo 2000 and Wasps loses 9/96 on time forfeit at TC 60s+100ms.jstanback wrote: ↑Sun Sep 15, 2019 12:09 amHi Ferdy,
Thanks for bringing this to my attention. Wasp handles the UCI_Elo command by reducing the nodes/second. At 1500 Elo the target search rate is only about 24 nodes/second. In my time handling routine I arbitrarily require the search to complete the first move of the 2nd iteration before it's allowed to time out. I think Wasp requires about 10 seconds/move at the 1500 Elo level to avoid losses on time and about 1 second/move at the 2000 Elo level. I'm running some games now at Game/60s + 0.6s with Elo set to 1900 with no time losses so far (after only 18 games), But that's probably the lowest Elo that will work with this time control. I'll see if I can fix this problem for the next release.
John
Code: Select all
# PLAYER : RATING ERROR POINTS PLAYED (%)
1 Cheng 4.39 UCI_Elo 2000 : 589 160 90.5 96 94.3
2 Amyan 1.72 UCI_Elo 2000 : 454 116 84.5 96 88.0
3 Stockfish 260819 UCI_Elo 2000 : 180 86 66.0 96 68.8
4 Rhetoric 1.4.3 UCI_Elo 2000 : 133 84 62.0 96 64.6
5 Discocheck 5.2 UCI_Elo 2000 : 127 84 61.5 96 64.1
6 Ufim v8.02 UCI_Elo 2000 : 116 86 60.5 96 63.0
7 Rodent IV 021 UCI_Elo 2000 : -7 80 49.0 96 51.0
8 MadChess 2.2 UCI_Elo 2000 : -132 81 37.0 96 38.5
9 DanaSah 7.9 EngineOpp UCI_Elo 2000 : -164 82 34.0 96 35.4
10 Wasp 3.75 UCI_Elo 2000 : -285 85 23.0 96 24.0
11 Arasan 21.3 UCI_Elo 2000 : -327 93 19.5 96 20.3
12 CT800 V1.34 UCI_Elo 2000 : -333 91 19.0 96 19.8
13 Deuterium v2019.2.67.71 UCI_Elo 2000 : -352 94 17.5 96 18.2