The monthly on-line engine blitz tourney for August will take place on:
Saturday August 19, 3pm Boston time (21:00 Amsterdam time)
To connect:
winboard -zp -ics -icshost winboard.nl -icshelper timeseal -fcp ENGINE.exe -fd ENGINEFOLDER -autoKibitz
(for UCI engines, add -fUCI)
On-line engine blitz tourney August
Moderators: hgm, Rebel, chrisw
-
- Posts: 27811
- Joined: Fri Mar 10, 2006 10:06 am
- Location: Amsterdam
- Full name: H G Muller
-
- Posts: 7220
- Joined: Mon May 27, 2013 10:31 am
Re: On-line engine blitz tourney August
What is exactly the difference between fairy-max and micro-max which is always playing in your tournament ?
-
- Posts: 27811
- Joined: Fri Mar 10, 2006 10:06 am
- Location: Amsterdam
- Full name: H G Muller
Re: On-line engine blitz tourney August
Micro-Max is hard-coded to play normal Chess. Fairy-Max started its life as a version of micro-Max that loaded its move-generator tables from a file, rather than having them as an initialized array. And I have a PGO optimized executable for micro-Max that runs 1.75 times as fast as compiles that I make myself (as I do for Fairy-Max). This is also the version tested for the CCRL rating list.
There are some subtle differences between micro-Max and Fairy-Max for normal Chess. The ability of Fairy-Max, to specify whether a move is a slide or a leap on a per-move basis (rather than deriving it from the piece type), and to also support hopper-type moves and moves that alter stride or direction after a number of steps requires some extra tests, slowing down the code a bit.
Also, the initial double push of Pawns is not hard-coded, but makes use of a general mechanism for specifying initial moves, and lame leaps. In the piece definition it is specified as a separate move. Where micro-Max always tried the single push, and when it succeeded and the Pawn was virgin, then tries to extend it to a double push, Fairy-Max first tries to do the double push, testing if the skipped square is empty, and tries the single push after that. This also seems less efficient, but actually it should be an advantage, because micro-/Fairy-Max is not able to perform a double push as hash move. (It would fail to set the e.p. rights if it does not come directly from the move generation.) For such unacceptable hash moves it would start move generation with the piece that made the move, however. So if the double-push is always the first move that is generated for a Pawn, it would still be searched first, even when not officially the hash move. This probably gains more in time-to-depth than the nps slowdown by testing for other moves will cost.
The latest versions of Fairy-Max are also a bit more stable, because they clear the hash table when null move is switched off. That way they cannot fall in the trap of trying to improve without null move on the depth of a search that was made with null move, after a hash hit. In micro-Max and older versions of Fairy-Max this could cause it to think for 5 minutes on a single move in a 1-min sudden-death game. Fairy-Max 5.0 is also able to recognize repetitions in search, while older versions only recognized repetitions of game positions.
None of this should cause a significant Elo change, however.
There are some subtle differences between micro-Max and Fairy-Max for normal Chess. The ability of Fairy-Max, to specify whether a move is a slide or a leap on a per-move basis (rather than deriving it from the piece type), and to also support hopper-type moves and moves that alter stride or direction after a number of steps requires some extra tests, slowing down the code a bit.
Also, the initial double push of Pawns is not hard-coded, but makes use of a general mechanism for specifying initial moves, and lame leaps. In the piece definition it is specified as a separate move. Where micro-Max always tried the single push, and when it succeeded and the Pawn was virgin, then tries to extend it to a double push, Fairy-Max first tries to do the double push, testing if the skipped square is empty, and tries the single push after that. This also seems less efficient, but actually it should be an advantage, because micro-/Fairy-Max is not able to perform a double push as hash move. (It would fail to set the e.p. rights if it does not come directly from the move generation.) For such unacceptable hash moves it would start move generation with the piece that made the move, however. So if the double-push is always the first move that is generated for a Pawn, it would still be searched first, even when not officially the hash move. This probably gains more in time-to-depth than the nps slowdown by testing for other moves will cost.
The latest versions of Fairy-Max are also a bit more stable, because they clear the hash table when null move is switched off. That way they cannot fall in the trap of trying to improve without null move on the depth of a search that was made with null move, after a hash hit. In micro-Max and older versions of Fairy-Max this could cause it to think for 5 minutes on a single move in a 1-min sudden-death game. Fairy-Max 5.0 is also able to recognize repetitions in search, while older versions only recognized repetitions of game positions.
None of this should cause a significant Elo change, however.
-
- Posts: 570
- Joined: Mon Jul 20, 2015 5:06 pm
Re: On-line engine blitz tourney August
Do you mean,hgm wrote:That way they cannot fall in the trap of trying to improve without null move on the depth of a search that was made with null move, after a hash hit.
"That way they cannot fall in the trap of trying to improve with null move on the depth of a search that was made without null move, after a hash hit."
-
- Posts: 27811
- Joined: Fri Mar 10, 2006 10:06 am
- Location: Amsterdam
- Full name: H G Muller
Re: On-line engine blitz tourney August
No, I meant what I said. What you say would be very easy, as searches with null move are a lot cheaper than without null move for the same depth. So it would just do the deeper search, and it would likely still take only a fraction of the time the original search that wrote the hash entry took.
The problem is that when in the nominal search time it gets for a move it reaches d=16, so that the position 2 ply later along the PV was searched with 14, all with null move on, that on its next game move, when the opponent played the expected move, it will get a d=14 hash hit in the root. So it will start iterating with d=15. But if the previous two ply traded away the last sliding material, null move will now be switched off, and searching d=15 without null move will take 10,000 times longer than searching d=16 with null move.
The problem is that when in the nominal search time it gets for a move it reaches d=16, so that the position 2 ply later along the PV was searched with 14, all with null move on, that on its next game move, when the opponent played the expected move, it will get a d=14 hash hit in the root. So it will start iterating with d=15. But if the previous two ply traded away the last sliding material, null move will now be switched off, and searching d=15 without null move will take 10,000 times longer than searching d=16 with null move.
-
- Posts: 4052
- Joined: Thu May 15, 2008 9:57 pm
- Location: Berlin, Germany
- Full name: Sven Schüle
Re: On-line engine blitz tourney August
hgm wrote:The monthly on-line engine blitz tourney for August will take place on:
Saturday August 19, 3pm Boston time (21:00 Amsterdam time)
To connect:
winboard -zp -ics -icshost winboard.nl -icshelper timeseal -fcp ENGINE.exe -fd ENGINEFOLDER -autoKibitz
(for UCI engines, add -fUCI)
This time Jumbo will not be present.
-
- Posts: 4889
- Joined: Thu Mar 09, 2006 6:34 am
- Location: Pen Argyl, Pennsylvania
Re: On-line engine blitz tourney August
hgm wrote:The monthly on-line engine blitz tourney for August will take place on:
Saturday August 19, 3pm Boston time (21:00 Amsterdam time)
To connect:
winboard -zp -ics -icshost winboard.nl -icshelper timeseal -fcp ENGINE.exe -fd ENGINEFOLDER -autoKibitz
(for UCI engines, add -fUCI)
dev-SF-McB is in - near final version version to be released early September ( could be final)
-
- Posts: 759
- Joined: Fri Jan 04, 2013 4:55 pm
- Location: Nice
-
- Posts: 7220
- Joined: Mon May 27, 2013 10:31 am
Re: On-line engine blitz tourney August
You don't get punished for that. If I don't show up in internal competition of my local chess club I lose rating points.
So you might as well disappear for twenty years and your engine rating stays the same.
Of course best is not to care about rating points for if you gain points you lose your precious time.
Sorry for wasting your time now.
So you might as well disappear for twenty years and your engine rating stays the same.
Of course best is not to care about rating points for if you gain points you lose your precious time.
Sorry for wasting your time now.
-
- Posts: 7220
- Joined: Mon May 27, 2013 10:31 am
Re: On-line engine blitz tourney August
Are there already sponsors found for your tournament?
Chess tournaments usually cost a lot of money they told me. Or is an online tournament comparable to bitcoin (less costs)
Chess tournaments usually cost a lot of money they told me. Or is an online tournament comparable to bitcoin (less costs)