I used to play in chess tournaments but have been gravitating more to backgammon. Today no human would have a chance against the top chess programs. I believe the World Backgammon Championship is a 25 point match. Is it still possible that a human could win against Gnu Backgammon or some other neural net program in a 25 point match?

If not what would the match have to be shortened to for a human to have a chance? I've really been curious about that for some time. Thanks for any insight.

Depends what you mean by "possible". With dice, anything technically could happen.

But realistically, the bot is a huge favorite, even against world class humans.

On the Elo scale, the best chess programs are about 500 pts higher than the top grandmasters. So you're right, in chess the best humans don't have a chance.

The Elo rating difference between Extreme Gammon, (the best bot) and the top humans is more like 75 points, so XG would be something like a 2-1 favorite in a 25-point match. I'm assuming that XG is playing at its XGR+ level, so it can play at a reasonable pace.

Thanks. I've never tried to play Gnu Backgammon a 25 point match because it completely destroys me on even 5 point matches and is very demoralizing. I used to think the program cheats but even throwing manual dice the results are the same.

One reason I was curious is I've never really heard of world class players going against computers except for the one game with Luigi Villa in 1979. For several years there was a lot of excitement when a chess grandmaster took on a top chess program but that ceased when all hope was lost for the humans. So I guess that backgammon programs are to that point now.

Thanks a lot. I was curious how much stronger bots are. So at least they can't crush humans in the same manner as chess bots. The unknown factor of the dice makes a big difference. Interesting.

This is interesting, I was thinking of a wider margin, more like 3:1 or even 4:1. Maybe 25 points is not so long a match as I thought.

I was thinking there is some sort of chart that would show the different probabilities in 25-point matches vs 7-point matches etc. I'm assuming the longer the matches the less luck is involved thus a big swing in probable wins and losses against different rated players.

A one point game would be almost like a crap shoot.

Basically yes, although I would say it a little differently. There is still just as much luck on each die roll. It's just that, over longer matches, good luck and bad has a greater tendency to even out - thus allowing the skill difference to show itself.

There is a formula and chart here.

In the late 1990s, Nack Ballard famously defeated Jellyfish, the first commercial neural-net program for backgammon, in a series of 300 independent "money" games.

Here is how Chuck Bower wrote it up:

A. Head-to-Head Money Play: JF vs. World-Class Humans

In 1997 Malcolm Davis initiated a contest by inviting two of the world's best human players, Nack Ballard and Mike Senkiewicz, to Texas to play against Jellyfish3.0. Human players put up their own money and Harvey Huie backed Jellyfish. Ballard and Senkiewicz were not teamed up, so actually there were two independent tests. Dice were human rolled to remove any concern that JF's generated dice were less than random. Each contest consisted of 300 independent money games. Coincidentally, Jellyfish finished dead even, beating Senkiewicz by 58 points and losing an identical amount to Ballard. JF's creator, statistician Fredrik Dahl, was quick to point out that a 58 point win in a 300 game sample is insufficient to conclude superiority. Ballard's win and Senkiewicz's loss were only significant at around one standard deviation each -- not particularly meaningful. Taken together, clearly neither the human race nor the droids could even hint at having an edge.

Source: http://www.bkgm.com/articles/GOL/Sep01/gol901.htm

Nack, who is retired from competitive backgammon, was recently elected to the USBGF American Backgammon Hall of Fame. He remains actively involved in backgammon study. He is considered by many, including me, to be one of the greatest backgammon players, analysts, and writers who ever lived. You can see his work, along with that of coauthor Paul Weaver, in Backgammon Openings, Book A.

I think Nack would readily concede today that XG, GnuBg, and Snowie can all outplay him.

Mike

How much better is XGR++ over XG+ in play and in analysis?

In tutor mode, I choose 4-ply for 1st pass, and XG++ for 2nd paSs if different. It's not that slow, but is it worth the extra time?

Mr Robertie,


500 points? even above Magnus Carlsen?

Is that why there has not been a "Deep Blue" type of competition in recent years?

I watched most of the games Magnus and the Gentleman from India play in Sochi in the world championship this year.

I used to play chess many years ago, never in a competitive environment and didn't play very good. However I do like to watch chess. I dont follow it closely tho, and was wondering way there has not been a "Human vs Computer" event (that i'm aware of) since Kasparov and Deep Blue.


Would it be a pointless endeavor these days? With a 500 Elo difference can a person win a game? Sounds pretty difficult, I have played 2000 Elo players a couple times and got crushed, I dont know what my Elo was, a friend said high 1500's perhaps 1600 on a good day. He was 1850 and i never won a game with him, I could have if I had a end game(but i dont at all)

I realize this is a OT question, i'm just curious.


thank you

Eric

Yes, such a match would be pointless now. Nobody disputes that the bots are better than the top humans.

Admittedly 500 elo points is pretty high. This maps to a winning chance around 95%. Can top humans really not hold the draw more often than one time in twenty? Even with white? Gosh.

The latest issue of Chess Life reports on the TCEC Tournament, a yearly even between the top chess engines in the world. This year's event was won by Komodo, which beat Stockfish 6 (last year's champion) in the finals by a three game margin.

Stockfish was rated 3329 last year, which means Komodo should be somewhat higher than that. Carlsen was rated about 2860 last I looked, so 500 points seems about right. At that rating disparity, Carsen would be expected to score about 2 draws every 10 games. Watching a match like that would have a certain ghoulish fascination, but I couldn't blame Carlsen for saying 'no'.

A bad BG player against a bot can also adopt a strategy akin to shoving blind every hand in NL. When in some terrible-but-not-hopeless position like a bad long race or a reasonably timed ace-point game, recubing every roll and trying to luckbox once for the match. I could start playing 99-point matches against a bot and I'd probably win at least one tonight with that basic plan. I could play the rest of my life against Stockfish and probably never draw a single game (without some version of playing the computer against itself by repeating lines enough), even though I'm relatively much better at chess.

Wow, I had no idea that the bots were that developed in chess. 3329 is just scary, 2 draws every 10 games. And thats last years bot.

I wonder where the ceiling is chess? I'm going to start following this, i'm not interested in playing chess. However the bot battle is interesting.


thank you for the information Mr Robertie,

best

eric

I wonder too. Most people I know agree that, with perfect play by both, chess is a draw. The fact that bots beat each other some of the time shows that there is still room to improve.

Not only would it be a draw every game (optimal bot vs itself), but it would be the same game (same moves) everytime!

which one is the best backgammon bot
XG better or gnubg?
and if you can tell me best setting. when i'm playing with grandmaster or xg r++ i won very easy its make many mistakes

I doubt that. There are probably many paths to a forced draw.

I am curious to see what you consider mistakes. Perhaps you can post some of them?

i take a sreenshots of its mistakes and then give it to you.

can you tell me best Player settings and General settings? an if you can give the the fotos of settings

Just use the default settings and you'll be fine. No need to overthink things.

I too would be curious to see some of XG's many mistakes. Sounds like you'll be able to teach many of us something about the game.

Sure, but the bot could use the same path over and over again. Depends how the bot is programmed.

If many paths have the same 'value' according to the bot, it should always choose the same arbitrary path (the 1st one for instance).

Unless the bot adds randomness to its algorithm, but I don't see many reasons. That could be:
1) If geared to play against humans, it could confuse them with a path they're less familiar with.
2) Just for the sake of not always repeating the same moves.

I guess it would be the programmer's choice.

Actually, this happens fairly often in gammon - for example, when all plays of a particular roll lead to double/pass. Do the bots always play the same move in this situation?

Judging by your screen shots, you have been trying to fine-tune things too much. Once you become more experienced, you can try setting up truncated rollouts that match your specifications, but for now I would go with the tried-and-true settings given below. Executing these commands will reset the things you have changed, which should be a good thing.

I recommend that you use nothing weaker then 2-ply, "World Class" settings. If you have a fast computer, you can try 3-ply, "Grandmaster." There are three places you need to make the change:

1. Settings > Analysis > Analysis Level: World Class
2. Settings > Players > Player 0 > GNU Backgammon: World Class
3. Settings > Rollouts > Play settings > First play both: World Class

You should also add a checkmark so the tutor uses the same settings as the analysis:

Settings > Analysis > Eval Hint/Tutor Level: Same as analysis

Is your version up to date? You can check the top line of the Build Info window to find out. I am running the latest Windows version: GNU Backgammon 1.04.000-mingw 20141021. Anything prior to version 1.00 is badly out of date. GnuBg introduced new neural nets in that version.

Help > About GNU Backgammon > Build Info

There is one important thing you must do to speed up GnuBg. Check to see that you have multithreading enabled.

Settings > Options > Other > Eval Threads:

A good choice is to enter the maximum number of simultaneous threads that can be supported in hardware by your CPU. For instance, if you have a duo-core Intel chip, where each core runs 2 threads, then you should set this to 4 (2 cores times 2 threads per core = 4). If you are not sure how many threads your computer supports, try looking up the CPU chip under the Intel heading at Wikipedia.

When you are ready to try a truncated rollout, check out the instructions for my XGR++ clone. They show how to simulate XGR++ in GnuBg. At the same time, they allow you to keep different settings when you want to do a full rollout.

Lastly, please include a GnuBgID when you post a board diagram here at TwoPlusTwo. In GnuBg, press Ctrl+C to make a copy of the ID, and then press Ctrl+V at TwoPlusTwo to paste it into a message.

Hope this helps.

Mike