Page 5 of 6
Re: Absolute ELO scale
Posted: Mon Dec 19, 2016 3:43 pm
by Guenther
hgm wrote:I don't think you can do that. There are plenty positions where repeating is the only way to save the draw. E.g. in KPK:
Say the above position results from a capture (Kxe3). Black's only chance to draw is take and keep opposition
1... Ke5! 2. Kf3 Kf5! 3. Ke3
This is not yet a repeat, because the previous time the white King was on e3 the black one was on e6, not on f5. But if black is now not allowed to play Ke5, because that will be a repetition, he has to play something else, after which white plays Ke4 and wins.
Well, but we are talking about the _worst_ player version.
It does not want to draw it always wants to lose.
Therefore it should avoid all repetitions, they just blow up the game length
against a random mover.
Re: Absolute ELO scale
Posted: Mon Dec 19, 2016 5:03 pm
by Vinvin
nionita wrote:Is there any theoretical problem if we define ELO 0 = play strenght of an engine which plays in every (legal) position one of the legal moves following a random uniform probability distribution?
The problem is that a lof of engines, even very weak, would score >99.999% against this random engine.
Re: Absolute ELO scale
Posted: Mon Dec 19, 2016 7:35 pm
by hgm
Guenther wrote:Well, but we are talking about the _worst_ player version.
Ah, sorry, I missed that. But when it selects moves with a low score, would't that make it automatically avoid moves that draw? I assume these would get a zero score.
Re: Absolute ELO scale
Posted: Mon Dec 19, 2016 8:06 pm
by Guenther
hgm wrote:Guenther wrote:Well, but we are talking about the _worst_ player version.
Ah, sorry, I missed that. But when it selects moves with a low score, would't that make it automatically avoid moves that draw? I assume these would get a zero score.
Yes. I thought so too, but in the example game a few posts ago there are dozens of two time repetitions. It is possible though that it was due to
being a game between two 'worst players' versions (and not worst vs. random as I thought first) and there was no worse move except repeating once?
May be those two time repetitions won't happen against a random mover, but I haven't checked it yet...
No idea how much time Daniel wants to spend on this ;)
Re: Absolute ELO scale
Posted: Mon Dec 19, 2016 10:43 pm
by cdani
Here they are:
Randscacs:
www.andscacs.com/randscacs089025.zip
Able to play arbitrarily long games, and also can play depth 1 random moves. Is updated with the current development version of Andscacs.
Andworst:
www.andscacs.com/andworst-0.3.zip
Able to play arbitrarily long games.
Re: Absolute ELO scale
Posted: Tue Dec 20, 2016 6:23 am
by Laskos
Great, Daniel! Works fine, I saw no disconnects. The longest game I saw was 1100+ moves in Cutechess-Cli.
Here are distributions of the game lengths:
And a reliable match result between A-worst and A-random:
Code: Select all
Score of A-Worst vs A-Random: 0 - 812 - 188 [0.094] 1000
ELO difference: -393.60 +/- 24.79
Finished match
The result is WLD
In A-Worst vs A-Worst most of the games (93%) were draws. So, we assist at draw death at both ends of the strength scale!
Re: Absolute ELO scale
Posted: Tue Dec 20, 2016 10:21 am
by Laskos
There seems to be no significant benefit from search beyond depth=1 for A-worst against A-random:
Code: Select all
Score of A-Worst depth=1 vs A-Random depth=1: 1 - 3162 - 837 [0.105] 4000
ELO difference: -372.48 +/- 11.69
Finished match
Score of A-Worst depth=7 vs A-Random depth=1: 0 - 3218 - 782 [0.098] 4000
ELO difference: -386.08 +/- 12.10
Finished match
Interesting, probably only a small benefit from depth=7. It's probably very hard to make use of search against random mover.
Re: Absolute ELO scale
Posted: Tue Dec 20, 2016 11:07 am
by cdani
Laskos wrote:There seems to be no significant benefit from search beyond depth=1 for A-worst against A-random:
Code: Select all
Score of A-Worst depth=1 vs A-Random depth=1: 1 - 3162 - 837 [0.105] 4000
ELO difference: -372.48 +/- 11.69
Finished match
Score of A-Worst depth=7 vs A-Random depth=1: 0 - 3218 - 782 [0.098] 4000
ELO difference: -386.08 +/- 12.10
Finished match
Interesting, probably only a small benefit from depth=7. It's probably very hard to make use of search against random mover.
The algorithm of A-Worst sure can be improved. I have done just very basic changes over the main Andscacs.
Re: Absolute ELO scale
Posted: Tue Dec 20, 2016 11:08 pm
by Uri Blass
Laskos wrote:Great, Daniel! Works fine, I saw no disconnects. The longest game I saw was 1100+ moves in Cutechess-Cli.
Here are distributions of the game lengths:
And a reliable match result between A-worst and A-random:
Code: Select all
Score of A-Worst vs A-Random: 0 - 812 - 188 [0.094] 1000
ELO difference: -393.60 +/- 24.79
Finished match
The result is WLD
In A-Worst vs A-Worst most of the games (93%) were draws. So, we assist at draw death at both ends of the strength scale!
If you want to lose the game the correct way is first to play well and later you have enough material to force the opponent to make checkmate against you.
I think that it should be easy to lose against A-worst by this tactics.
Re: Absolute ELO scale
Posted: Tue Dec 20, 2016 11:49 pm
by Guenther
Uri Blass wrote:
If you want to lose the game the correct way is first to play well and later you have enough material to force the opponent to make checkmate against you.
I think that it should be easy to lose against A-worst by this tactics.
This is right for losers chess and I also thought about this before, but I think in chess (different rules - no recapturing is required!) it's not that easy as it sounds.
I am tempted to try to lose against 'Andscacs worst' tomorrow, may be you can also try an example?