Not logged inRybka Chess Community Forum
Up Topic Rybka Support & Discussion / Aquarium / Thoughts, ideas, and request for comments
- - By centipawn (*) Date 2021-01-12 11:22 Upvotes 1
Hi all,

after a little over one year of playing on ICCF, I want to share a few thoughts, theories, and questions. My main objective in doing this is to get a discussion going, but also maybe to get some insight into things I am missing and should be looking into.

So, I have played on ICCF for just over a year now. I have won some games, drew some games, lost none. As my rating increased and my opponents grew stronger, it has become harder to win, but I still seem to win some games. I use Aquarium, and have created an opening tree covering almost 5 million positions now. When a new game starts, I use this tree to make my decisions (and do some research in about 80 pounds of chess books I own, sometimes adding lines to my opening tree), and once it gets to a position where I have an N less than a few thousands in that tree, I will fork it out into its own IDEA project.

From there, I will generally, for each move, run infinite analysis and look at the top three moves.
- If it's obvious that there is one good one and the other two are much worse, I will make a decision very quickly (often just letting IA run for 5 minutes) and make the good move.
- Otherwise, I will let IA continue for up to 3 hours. If it stabilizes and there is a clear favorite, I will make that move.
- If IA says I am winning, I will make the top move.
- If IA says this is a dead 0.00 game, I will let it run for a while just to be sure, then make the top move - and sometimes, will go to IDEA as per the next point regardless.
- Otherwise, I will let IDEA do its thing and then decide, sometimes also adding moves I think could be good, and/or adding IA for such moves as well. The most difficult cases for me are those where IDEA reports, after considerable time, a number of perfectly equivalent moves (also equivalent in the diagrams). Sometimes this is because move order doesn't matter in those cases and all those moves will be made anyway in some order - that's easy to resolve. When this is not the case, I try to make the move that I feel keeps up pressure and/or adds more danger to the opposing side (which is probably just a figment of my imagination - otherwise, IDEA surely would express that).

This has served me well and I have not been in serious trouble so far even against SIMs and GMs, which is why I am increasingly beginning to wonder how others sometimes end up in trouble, i.e. losing. From what I have seen, factors appear to be
- Blunders when entering moves - e.g. playing Nd5 instead of Bd5 at some point where Nd5 is the obviously better move.
- I think some players may occasionally feel there is only one good move and make that without checking properly.
- Some do not use latest / current engines.
- And some may not let their engines crank away long enough.

Of course, at some point, I am bound to fall into one, or a few, or all of these pitfalls myself. But for those with good hardware, time management and diligence, a draw seems very nearly unavoidable. There are more than a few players on ICCF who have been playing for a while, in some cases in many many events in a short time, and have never lost a game so far.

All this caused me to spend considerable time thinking about how to win against such players. Essentially, the plan would be to identify positions where strong engines need considerable depth / time to avoid making a suboptimal move. That is theoretically not so difficult to do - simply by letting SF run to depth 50 and figuring out for which positions the best move at depth 50 is different from the one at depth 48, for example (exact depths and parameters would have to be chosen depending on opponent strength etc., but I'm sure you get the idea). In short - find those positions where the potential for mishaps is biggest. This is not really different to using IDEA, and it feels this is essentially equivalent to investing more CPU time than your opponent.

Has it all just become a question of how much CPU power you invest then? Am I missing something? Do SIMs/GMs have tricks I am missing here, beyond possibly a really good OTB understanding and intuition of chess?

As an additional side note: I believe Aquarium as a software product is dead, and just carried over from year to year to get as much money out of a dead horse as possible. I have written my own tooling (which essentially does the same as IDEA), but as that has no GUI worth mentioning and produces the same results as Aquarium, I am still using Aquarium mostly for now. However, if anyone does have any ideas how to get better other than through means that can be achieved by running IA and/or IDEA for longer and/or to greater depth, then the skills to make this happen do exist (probably in many others as well as myself).

So - in the name of all 32 pieces, how does one best go about beating someone with plenty of CPU power, IDEA and the latest engines?!

Thanks for reading through all this. Feedback highly appreciated!

Parent - - By pawnslinger (****) Date 2021-01-12 16:42
"So - in the name of all 32 pieces, how does one best go about beating someone with plenty of CPU power, IDEA and the latest engines?!"

This is the big question.  I guess if we knew the answer, we'd all be GMs by now. 

My IDeA trees are so big now that I seldom use anything other than IA and of course the trees for reference material.  I too have a large stack of Chess books, but find them pretty useless in the Stockfish age.

Most of my wins come from 2 factors... my opponent blunders or the opening has led to double edged positions, where even computers can go wrong.  But those positions are increasingly hard to find.

I find computers look for draws too often, so I have gone back to using the "contempt" setting... and increased the value by about 2X.  This tends to help the computer look at riskier positions.  Often I will make a provocative move (not the computer's top choice) to encourage risky play.

And even with all I do, I seldom lose or win a game.  This has caused my elo rating to gradually taper off.  So I think I have reached my high point.  Which was around 2450... I hope I can keep it above 2400, but that is an open question.
Parent - By braindied (**) Date 2021-01-14 08:43
"I guess if we knew the answer, we'd all be GMs by now." coupled with the observation in the same thread that 220/225 games between higher rated players ended in draws, means that becoming a GM is getting harder and harder regardless of playing techniques.

Something like a Category 11 event with a +3 score requirement is very rare now to achieve. And finding a Cat 12 or above event is becoming very hard - even the latest World champ final is only cat 11.

Coupled with the trend towards blitz speed playing at correspondence chess - the lifespan of true CC play must be limited. No wonder the Aquarium developers probably feel they have captured all the market they can, and have moved on to more lucrative things.

Looking at my achievements tab on the ICCF server I was surprised to see I have a 91/494 games win ratio. But since 2016, a +3=86 record. In 960 chess a +15=23-0 record. No wonder 960 looks more appealing.
Parent - - By dickie (**) Date 2021-01-13 18:08
Sadly there is not a great deal of life left in CC, at least in standard chess. Stockfish is just too strong now, even on modest hardware, and can easily defend positions that were once felt to be promising. Looking at the ICCF archive for 2020, of the 225 games played between players rated 2500 and above, 220 were draws. I have 7 GM norms and yet have not won one of fifty games in the past 4 years. I am afraid there is no ‘GM magic’, just the lottery now of finding the rare opponent who might blunder or make a clerical error.

As for Aquarium, I think you might be able to make better use of it. I find adding tree configurations that display modern human theory are much better than engine evaluations for guiding the opening. In the middle game I use IDeA mostly in automatic mode. If you spend a little time calibrating IDeA’s settings then the trees should be more informative than plain IA. A good way to calibrate the settings is to build lots of small trees using Leela at 1 sec or 1 ply. It takes only 2 or 3 minutes to complete a 5000 task tree and the evaluations are pretty sound. Begin with the root at the start position and then perhaps try a6 of the Najdorf and adjust the variables, tree shape, tree width, number of alternatives and project score bounds until you find a combination that builds trees to reasonably reflect human theory in both width and depth. Other than that I have found the Branching Factor and Best Line functions quite useful at times. IA is more useful for checking for tactical oversights, or once the outcome of a game has become clear.

Finally, with IDeA I find it is best to turn off Hyperthreading or Simultaneous Multithreading. In fact I would go as far as to say that IDeA is unusable with either turned on. Too many rogue evaluations get sent to the tree. Turned off, this problem disappears.
Parent - By pawnslinger (****) Date 2021-01-14 17:19
I have thought, for a long time, that ICCF was making a mistake with their event structures.

Something as simple as a "ladder game" would help a lot... in my opinion.  Something that would allow lower rateds to play higher rateds.  With players being confined into stratified layers, there is little natural inclination to risk elo by playing outside one's layer.  This is true, in my opinion, and will lead to a large number of draws, as we play other players of equal strength.  Computer playing programs only makes this worse.  As human blunders are far less frequent.
Parent - By centipawn (*) Date 2021-01-14 06:52
Thank you both for chiming in... and yes, of course there is no clear answer to the provocative question I asked - otherwise those in the know would have the same problem all over again when playing against each other...

My chess books are not entirely useless. They provide a little amusement when analysis proves the author is wrong (Mar del Plata...), but more importantly, I add sidelines from the books to my trees sometimes, thinking it may help IDEA not to miss something.

I will give the contempt setting another chance. I have not played much with it so far.

As for the suggestion to use short time controls to build opening trees, I will also look into this. I have primed my trees from the ICCF archive so far (from all games where both opponent have a rating > 2350, the first 30 moves are all in my tree... that covers quite a lot), so it will be interesting to see if STC engine play can add anything valuable. I have tried various settings (branching factor etc.) in IDEA, but believe that it does not make much of a difference in the long run.

Hyperthreading is on for me, and I have not found any problems with rogue evaluations so far. Could be that I just haven't realized yet, but it seems unlikely.

Also - "at least in standard chess" - yes, I have played one Chess 960 tournament so far and have found much more potential for mischief there. Maybe this is the way to go - but it seems not many people on ICCF play Chess 960 so far.
Up Topic Rybka Support & Discussion / Aquarium / Thoughts, ideas, and request for comments

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill