Asymmetric match equity tables for gnu?

[blitzxz]

Where can I find these? There are only two of them as default and I can't find the rest xml-files from anywhere. I could make the tables myself but I don't know how to do it (propably with some xml-editor?) and I don't have the correct values either. Help would be appriciated..

[playBunny]

These might help if you need to take the roll-your-own route.

How to Compute a Match Equity Table  ..  http://www.bkgm.com/articles/met.html

Sho's Backgammon pages  ..  http://www46.pair.com/sengoku/index.html

[blitzxz]

Well, that didn't help much. I would need the "fish tables" by Walter Trice, which I still can't find in easy format (numbers in a table) online. I also would recommend everybody to try to analyze your games with these tables. In 5 point matches even the 100 rating point difference can change just about everything when there are lots of gammons. And now I'm sure that doubling in match play should be always done according to opponent. It would be a interesting test to switch this 100 points better table to a fibs bot. It should improve it's winning chances.

[dorbel]

Altering your doubling strategy according to the difference between your rating and that of your opponent comes up against a basic flaw, which is that the ELO system in use on fibs (and elsewhere) is far too unstable to be useful. An individual's rating range quite normally covers 200 points and 300 points is not unknown. If your chess opponent is rated at 2200 (say), you have a fair idea of how well he can play. An opponent rated 1800 on fibs maybe a player at the bottom of an 1800-2000 range or the top of 1600-1800! The fibsleagammon ratings are more stable for several reasons, (no selection of oponents, no droppers, all matches of a meaningful length and most matches against players who are roughly of your own strength) but even these see variations of 150 points quite normally.
What you need to know is how well somebody can play the checkers in the type of game under consideration and what their reaction is likely to be in response to your cube action. You will, IMO, benefit far more from remembering who can be relied upon to butcher a difficult positional struggle and who will never recube to 4 until it is a pass (for example), than you will from trying to construct assymetric match equity tables. Quite apart from anything else, remembering a normal MET is hard enough without having to then adjust it!
By all means buy  "Can A Fish Taste Twice As Good?" by Trice and Jacobs and the disc that accompanies it, but don't expect it to up your win rate much.

[blitzxz]

Remembering match equity tables is not what I'm here after. I don't work doubling decision that way at all. I remember lots of reference positions. Calculating take points is a too slow and still unreliable because you have to guess gammon rates and winning rates which is even hard for bots. Analyzing with asymmetric tables I get the general feeling how you should play against different opponents. And of course I'm trying to consider opponents reactions to cube actions or his big checker play errors. However this is as hazardous as guessing opponent true ratings and if you're using normal match equity tables you're still guessing every one is equal. In general doubling is very hazardous business especially in match play and largely influenced how you're opponent is playing. The theoretical way is of course to expect that opponent and you are playing perfectly but I don't think that's a good starting point to any game that is not very simple. The question here is: Are you playing against snowie error rate or are you playing against your opponent? I prefer the later.

[Zorba]

I don't know of any existing METs outside the ones in the already GNUBG package. The xml files have a pretty simple structure, load one in Notepad and you might be able to identify the structure of it even without any explanation. So if you have the numbers of other METs, it's not hard to make them into a xml file. You can email me for help if you like (I created the g11 table some years ago).

When using asymmetrical METs for analysis, you have to be careful how to interpret results. There will be an inconsistency, because the MET assumes unequal strength (for the rest of the match), but an evaluation of w/g/bg chances is still based on GNUBG's neural nets assuming equal play. This can lead to very strange results, for instance in free-drop situations.

An example: suppose the MET assumes you're a 55-45% favorite in a 1pt match. The score is 1-away 2-away, Post-Crawford. You won the opening roll with a 6-5, 24/13. Your weak opponent makes the automatic Post-Crawford double. Should you take?

Of course you should, starting the game with a 6-5 is pretty strong. You're 53.3% to win according to GNUBG 2-ply. However, using the asymmetric MET, GNUBG will say "pass", since it compares this 53.3% to the 55% you're supposed to have in the next game. The flaw is clear: since you're 55% to win a game before the opening roll, you should have even more after an opening 6-5. But that part is not reflected in GNUBG's evaluation.

This is true for all sorts of positions and it often has a balancing effect on doubling decisions. A strong player should tend to double later if you just look at asymmetric METs, but since the stronger player will also get higher winning chances from the current position than GNUBG indicates, the stronger player should double earlier. Overall, it tends to balance out.

Not always though: some positions require little skill from one or both sides (race/bearoff), some positions are very hard for weak players (backgames/primes), etc. There it can really pay off to change your cubing strategies according to opponent's skill level.

When using GNUBG though, you'll always have to ask yourself how its w/g/bg percentages should be adapted to reflect different skill levels, not an easy task.

[blitzxz]

This is true for all sorts of positions and it often has a balancing effect on doubling decisions. A strong player should tend to double later if you just look at asymmetric METs, but since the stronger player will also get higher winning chances from the current position than GNUBG indicates, the stronger player should double earlier. Overall, it tends to balance out.

Good point. I didn't think about this through.

[dorbel]

Good writing by Zorba here, a difficult subject approached with perception and clarity. Thank you.