Not logged inRybka Chess Community Forum
Up Topic Rybka Support & Discussion / Rybka Discussion / Contempt question
- By Thin Ice (**) Date 2009-10-02 10:26 Edited 2009-10-02 10:30
If one assumes that a correspondence opponent is also using RYB 3(or another top notch chess program) like me with long time controls,is it best to use a zero contempt value?
Or is it better to simply use default 15 all the time?What about a different scenario....where we are entering mid game,and i'm using default 15,we are "even"in piece value on the board,but positionally,Ryb shows a fractional disadvantage for me,say -.20.At this point might it be better to switch from contempt default 15,to contempt 0?
Just asking,because this contempt value stuff still holds a mystery to me.
- - By Eelco de Groot (***) Date 2009-10-02 13:49 Edited 2009-10-02 13:52
As far as I understand it, if you are using Rybka in analysis mode, the contempt value is automatically set at zero. You can look that up in the Rybka FAQ. I don't know much about things like IDEA but I would presume there too you are doing analysis so contempt is not used, set at zero. For playing correspondence chess I think it is generally better to expect much resistance from your opponent so again better with a neutral viewpoint.

More in general, from a programming viewpoint, you could say something more about 'contempt', and about what happens in the searchtree if the eval is close to a draw value, but if that is too technical for your purposes please ignore the rest of my post :)

Contempt is more complex in Rybka but for most programs you can simply change the evaluation of what a draw is worth, that is the original interpretation of 'contempt'.

The way I understand it, Bob Hyatt has posted about this in the past on CCC and he could probably explain it much better, but my view of this after reading mainly Prof Hyatt's comments is that it is often just not safe to use a value different from zero. And if just one program says that something is exactly a draw you should not trust it too much. Programs already have to interpret the position for both sides, if they interpret it better for one side and they are wrong, there is not enough compensation in the lines in its searchtree by having a real opponent that may still continue playing for a win, the program automatically thinks if the position is better for him, it must be worse for the opponent. This is the epitomy of machine logic of course but a human would not necessarily think that way! Whenever the position seems to get worse for one of the sides in the alpha beta tree, this side will try to gain a draw only. But if the program gets it wrong, the 0.00 score that you see in the end is wrong. You would maybe get more accurate results if both sides held a slightly different view in the analysis tree and kept playing longer for a win even if they thought the position was worse for them. But it is dangerous to just change the value of the draw, in the case of Rybka contempt is more complicated than that I am sure but it would not be a surprise that you can see some of the same type of problems there.

Eelco
Parent - - By Uri Blass (*****) Date 2009-10-02 14:02
The default contempt value in analysis mode is zero but you can change it so nothing is automatic.

The contempt playing value has no effect on the analysis but the contempt analyze value can change the analysis.
positive numbers of contempt analyze mean that white tries to win the game when negative values mean that black tries to win.
Parent - - By Thin Ice (**) Date 2009-10-02 14:38
Thanks for the input.Buts lets say for example that I am NOT using analysis mode or infinite mode,but a standard time control setting of 40/2 with time constraints.I do this to try and make RYB make its best move in a shorter amt of time since i do not have the desire or time to let Ryb analyze for hours and hours on each move.Then,would fooling around with this contempt be wise?Going from +15 which is default,to say 0,depending upon the position or disadvantage on my side?
Parent - - By Uly (Gold) Date 2009-10-02 17:51
Set it to 0, at correspondence games you don't want to use negative contempt* as that will lead to draw on won positions and positive contempt* will lead to overreach trying to win drawn positions. Of course it's more important to force the moves and check the variations, sometimes that 0.40 you're seeing on your main line may fall to 0.00 when you force the moves at same relative depth (which means you can see the fall immediately), etc. so interaction can be critical and sitting at the root is not recommended (unless you're planning to interact later on.)

*) relative to the playing side.
Parent - By Eelco de Groot (***) Date 2009-10-02 23:36 Edited 2009-10-02 23:42
Interesting about Joey's approach is that it would tackle some of those possible problems that I mentioned, that if the two sides both have the exact same opinion they will 'cooperate' towards a result as it were, especially around zero, and there you sometimes would like to have slightly different opponents, one attacking more and the other defending. A similar thing can be achieved by letting two programs with different playing styles play in your 'shoot-outs', Monte Carlo or trial games etc., that is usually better rather than just Rybka against Rybka, and in this case there is a human element by choosing how much one program plays for a win and the other for a draw.

It is related to what Uri mentioned that you can have Rybka analyze with different values of contempt, if you think that White is better maybe contempt should be raised, if you think Black is better and you are playing White yourself then you can set it lower to see if by playing/analyzing for a draw the analysis improves, from your point of view.

In searchtrees, breaking the symmetry I suspect will at the least lead to bigger searchtrees so there it is even more of a question how you would go about this, and more of a question if you can get better results by having asymmetric evaluations around the draw. In general it is usually a recipe for disaster to change the evaluation depending on the circumstances and I certainly would rather have evaluation as a constant. I'm really speculating here, maybe Vas would have a different opinion based on his programming approach and the results in Rybka with 'contempt'.
Parent - By Rowlando (***) Date 2009-10-04 08:01
Your analysis method will have a much greater negative impact on your results than anything you do with the contempt setting.  Auto-play is asking for trouble - if the big fail events don't kill you, the razor sharp positions almost certainly will.  And if your opponents ever figure out what you're up to, look out....
Parent - - By Geomusic (*****) Date 2009-10-07 08:13
i never understood the point of correspondence chess after the advent of master strength chess programs. Why bother, if I wanted to play a (C) I would play my own desktop.
Parent - - By Uly (Gold) Date 2009-10-07 08:37

> if I wanted to play a (C) I would play my own desktop


But you can't download the human, Corr chess is not only (C)
Parent - - By Vempele (Silver) Date 2009-10-07 08:46
Besides, correspondence chess is much more efficient - you get to use your opponents' computers for free!
Parent - By Uly (Gold) Date 2009-10-07 19:17
Was that the point of the Frakula game? :yell:
Parent - By buffos (Silver) Date 2009-10-07 08:45
do you think that CC chess is just computer moves? If you rely on your pc, you will just plainly lose. You should invest many hours of analysis, PLUS understanding is very crusial, since evaluation alone says nothing. Its just a tool.
You could spend days in finding a good plan
- - By Thin Ice (**) Date 2009-10-05 08:47 Edited 2009-10-05 08:57
I'm not convinced that using analysis mode over using a 3 stage time setting is advantageous,atleast for the most part when playing alot of corresp.games at one time.I have been winning most of my corresp.games and had a bunch of draws,but only 2 losses so far out of 102 games and both of those losses were black and a style response that I chose,not Ryb which was a semi slav meran.
When my clock runs too far down,I just add time,overclock or even reset the pieces up from that mid or endgame position.I have many times compared the infinite analysis response to my 300'/40+120'/20+30' setting(for example)and they were the same Ryb move response,closed or open positions.I'm sure there are exceptions,but in the long run,if i'm playing 30 or 40 games at one time(which is common among corresp.players,or even 10 games at one time),I do not have enough hours in the day to set those games on slow analysis mode.So the overall results in playing "bulk"amt.of games using Ryb and time settings like I described above still averages out on an excellent scale,even if I do get an occassional loss...and my corresp rating still goes up and up!
I did notice to get to the same"depth" takes much longer on analysis than my 300'/40 setting(even using my monster quad) and generally the move response is the same...most of the time,atleast 98-99% of the time.
I do fool around with dropping from default +15 to default 0 when I see the game is extremely drawish and that has changed more of a move for me from time to time than analysis mode....and usually to my advantage.Other than that,I don't care what the Ryb experts say about analysis vs time control settings,I have seen very little difference in the move outcome on my side...atleast in the long run.When you stick with only the highest stat. openings,have 300' or slightly more time settings and have a quad,the advantage of using infinite analysis over other settings is negligible at best.
Parent - Date 2009-10-05 18:38
Parent - - By Permanent Brain (*****) Date 2009-10-06 01:06 Edited 2009-10-06 01:10

> I did notice to get to the same"depth" takes much longer on analysis than my 300'/40 setting


That may point to a wrong setting. Basically, an engine analyses with the same speed in game mode and in analysis mode (single variation). But note that the time to reach a certain depth in analysis mode is increased, if the number of variation in multi-pv (multivariation mode) is increased. Not 1:1, in other words 2 vars don't require twice the time, but it is somewhat slower. I think up to 3 or max. 4 variations are useful because you see e.g. if alternatives are interesting to consider. I use 2 for "easy" analyses, but I don't play corr. But more that 4 I would not recommend, because who is interested in the 5th best, 6th best etc. and then the slowdown will get bigger.

> and my corresp rating still goes up and up!


I am afraid this will end if you reach higher rating levels, where probably more powerful computers, several engines and more complex concepts are being used than among the opponents you beat now. Because if a simple concept of "only Rybka (as it seems), and simple straightforward calculation like in a engine versus engine game", has significant success in top segments, then I must lose any respect in high rated corr. players. :lol: But I guess that will not happen.

P.S. Do you play in the ICCF, or in another corr. organization?
Parent - By Geomusic (*****) Date 2009-10-08 11:32
i bet if they took a statistical avg of all CC players ratings both CC and classical the differences would be several hundred pnts per chess player, sad...
Parent - By Rowlando (***) Date 2009-10-06 19:17
The sharper the game, the more likely you are to suffer as a result of your analysis method.  I suspect that there are a lot of players who use the same approach as you which is why I always steer my games into the sharpest positions possible.  The more fail events, the easier it is to win.  If it works for you by all means continue it but just know that you're going to eventually reach a rating ceiling as you start to encounter stronger and stronger opponents.  The top echelon of players will consistently defeat you from both sides of the board -- eventually you'll have to start using a different method if you hope to compete with them. 

If I knew beforehand that my opponent was going to use this method in a game with corr time controls, I could run the table.
Up Topic Rybka Support & Discussion / Rybka Discussion / Contempt question

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill