Using engines to create a database: what ideal time control

[ozziejoe]

Hi folks,

I've set up a book that plays only my openings. My intention is to build a database, so that i can look at some good opening lines to try. I'll be able to build a bigger database than is avalable in my hugebase (although the human database has 5 million games,it often has only a handful of games at gm level in my chosen line).

So.....What do you think is the best time control, which balances the need for quality with the need for speed (e.g., so you can tell if a line has a reasonable winning percentage).

I am leaning towards 4 min 2 sec games. What do you think?

I know longer time controls are considered better, but better for what? I think longer time controls have two disadvantages

1) they don't allow you to generate enough games in a reasonable period of time, to get a reliable winning percentage for a line

2) they may cut down on variety, as engines with longer time may gravitate towards a best move. I would like to see lines which might be slighly inferior (say by 5 to 10 centipawns) but make sense to me and give practical chances.

So please help me with some suggestions.

[Eelco de Groot]

Hello Joseph,

If I were to do such a thing I think I would find out if Romichess could play with your book on Playchess so it could learn against and from all the different opening books out there. Playing games against an engine that has no more bookknowledge than what is already in your book will not be very efficient I think...

Regards, Eelco

Edit: I see from Michael's answer in thread below that this is not yet possible. You could let Romi play with no starting book and then see what lines it develops but it is not very likely it will follow the lines you are interested in...

[ozziejoe]

Yes, romi is another option I am exploring, but i need to find a way to combine her learning with an opening book. one possibility (if the book is highly selective) is to take key positions from the book at let romi train on those. It was learning ok when i loaded a pgn, rather than running things via an opening book

best
J

[Ovyron]

I am leaning towards 4 min 2 sec games. What do you think?

I think that 2 seconds is a very short increment, because when the games go for too long, the engines have to start playing at 2 seconds per move, and sometimes that's just too short and the games results are more random (An engine loses a won position due to having to play too fast.)

That's why I support long increment time controls, like 1'+5" (It gives the same game length than 4'+2" for a 60 move game, with the advantage that the engines get 5 seconds per move in long games, instead of 2 - That's more than twice the allowed time, and could get decisive on game results so I think it's worth it so the game results are less random.)

[ozziejoe]

ok, that makes sense. Engines are often a little weaker in the endgame, and we don't want to give them too little time.

i'll go to 4 minutes with a 5 sec increment.

best
J

[Dirt]

ok, that makes sense. Engines are often a little weaker in the endgame, and we don't want to give them too little time.

i'll go to 4 minutes with a 5 sec increment.

best
J

I would suggest varying the time somewhat, say 1/3 each 3, 4 and 5 minutes. This should lead to more lines being considered.

[Henrik Dinesen]

Having done/doing some the same myself, I've found - not surprisingly - that there are different usefull approaches:

1) Having games played with a fixed TC like this thread is about
2) Setting up a TC that gives a lot of depth in the first X number of moves, since some lines is handled badly by engines in general; due to the longterm perspective in the line(s) more time is needed to avoid all kinds of premature moves - and then continue the rest of the games with a lot faster TC, again weightning the kind of game...
3) Use of the Rybka randomizer, varying the randomize setting
4) Romi as suggested - haven't startet to use it yet, but the attempt is on the "important to try list"

Normally I make a sample of positions with my candidate moves, my replies to a large number of moves.
I make the postions from either from a book made from games with only the opening in question in it, and a database with my own keys only following my prefered moves for the side I'm playing, and "everything" for the opponents side. Of course I skip lines when they've been proved bad enough for further investigation, but surely a lot of odd looking moves would be nasty surprises OTB, and is handled well by engines, so I like to have them in my keys.
Also here, I tend to look at moves that I may fear and ad them.

Bottomline, I don't use simple TC's before I'm conviced the postion is stable enough for the engines to make something decent out of it.
Howver, the approach with simple TC's will after a lot of games give a clue about if an engine is better in handling some kinds positions, nice to know when the analysis starts