Exactly.

I don't know, I assume they don't always do. When I play against XG Mobile on my phone, it doesn't always open 6-5 with 24/13 for instance.

Regarding different plays that lead to a double/pass, I guess it doesn't really matter because the bot will have to pass no matter how it plays. But again, programmer's choice in this case. Maybe the progammer wants some variety or wants the bot to appear more "human".

Given the dice factor does anyone think that computers have reached their peak against humans which is about 75 elo points? I'm still a little surprised that there are not more top human vs computer matches in backgammon because of the press and buzz chess programs vs humans had for several years.

1. chess is way more popular to begin with
2. in chess, variance is almost nonexistent, so skill-end result correlation is easily understood by a general public whereas in backgammon..
3. ..general public largely thinks bots are cheating
4. limited popularity of backgammon leads to no big corporate sponsors/promoters to advertise a match like this

What about bots topping out in comparison to human ratings?

You can set the level higher and stop that nonsense. Unless you like it of course.

the highest setting for extreme gammon plays at 0.11 pr (performance rating, 0 is a perfect score) as compared to rollouts, so even if backgammon is not entirely solved by it, it's pretty close

humans obviously have a much bigger room to improve, the very best players average around 3pr iirc

Excellent point sir.

there are a couple players now who are a hair below 3pr, there are still getting their pr lower over the past year as they are improving. amazing. i had this conversation with a less than 4 pr player last night. we both agreed on our observations, i just mentioned.

Gammon is almost entirely luck you can be great grandmaster and still get beaten by a noob. It is a myth that backgammon requires skill to be good or to win games, it is down to Luck, glorified snakes and ladders. so no the game is just like snap or any other casual game, it isn't like poker at all since you can't see their cards, so that requires more logic and math/probability. this is just hurrr i roll dice and move there. I am an awesome backgammon player who would outplay pretty much anyone but it doesn't matter since it is so much luck involved.

How can you be an "awesome backgammon player" if there is no skill involved? If it is just roll and move, then nobody can be any better than anybody else.

We actually haven't had a good troll on the site for quite a while. I was wondering where they all went.

If by good troll you mean obvious and not subtle at all, then yes, he's good.

Subtlety and nuance usually aren't a troll's strong suits.

How much better does XG play at its slowest (most recursive) setting than at XGR+?
How much of that 75 point differential is a function of: fewer oversight errors; raw match equity table crunching; perfect post-contact play...? I.e: do you believe that a human can train and learn to make up that differential, or will we always be the frog jumping halfway out of the well...

Corollary: can humans invent a board or card game with objective win/loss parameters that a bot will never be able to beat?

And all the oxygen atoms in a room might wind up in a tiny corner for a split-second...

Yeah, your win% may be >0- unlike at chess- but it will be fast approaching. Of course, that doesn't mean backgammon is easier to master than chess. Not by a ...long shot. You don't need to crunch numbers at chess- you just rollout moves in space. Sure, you must learn themes and strategies to beat other humans, but the bots find chess rollouts routine these days.

Maybe the reason XG can't crush Robertie et al so easily is that the numbers get so big so fast that the themes remain more...essential... posing problems a neural net can't hope to process even at nanospeeds...

They tried: I believe "Game of the Amazons" was designed to be difficult for computers. Given the tree search stuff from earlier times this was somehow successful.

Having the image based learning stuff (i.e. AlphaGo and it's cousins or was it Warcraft(?), a real arcade computer game) I doubt that it will be possible nowadays.

best
Frank

Simple answer to the question of humans beating the bots. Of course they can and do beat them. Chess and backgammon are entirely different animals. Magnus Carlson will beat you blindfolded. Mochy will probably beat you but if you are a good player(PR in the expert range or better) then you can beat him with reasonably good dice. Yes, backgammon is a game of skill but it is also very much a game of luck and a gambling game. The exact percentage of skill to luck is debatable but luck very much exists. There is no luck in chess.

"Of course they can and do beat them."

Not any more.

If you read above, the question is whether a human can ever regain the edge in ELO rating versus a bot like XG. Humans lasted longer at BG than Kasparov v. Big Blue.. but those days are gone. The bot plays perfectly, I believe, after contact. A human might get there with a ton of training, but that is not the main equity leak. The bot actually factors "luck"- and although that algorithm may be somewhat approximate, the ELO rating is only advanced with hundreds and then thousands of iterations. So no, no human alive now can outrank XG...for which host grandmaster Bill Robertie is a reliable source.

I am looking for similar help. I have a fast server for my extreme gammon 2 to play upon. I have two questions:
1) How do I get the strongest level of opponent short of playing against rollouts? My extreme gammon ELO is about 2070 or so says XG2.
2) Why use variance reduction when doing a rollout? Isn't that just a slow way to get/guarantee a nearly identical answer? My reasoning is that it is the NN that decides what values/weights to assign to every roll when using variance reduction. I currently am trying truncating the rollout after an 8 ply depth doing a 3 ply search using xgroller+ with gigantic to minimize the effect of the NN. XG2 is awesome except when it doesn't understand the position which is the only thing that makes sense to me to rollout.

Help/thoughts will be appreciated!
CT

I love "Game of the Amazons." (It's also just known as Amazons.)

But no, Amazons wasn't designed to be difficult for computers. I've never read that about Amazons, and if it was, someone did a very poor job. For example, there are only four pieces each and each piece has the same powers. With each move, there are fewer squares on the board that can be moved to. The rules are simple and strong bots already exist.

I strongly suspect you're thinking of the board game Arimaa. According to wikipedia, Arimaa is a two-player strategy board game that was designed to be playable with a standard chess set and difficult for computers while still being easy to learn and fun to play for humans.

Playing against the default XGR+ is fine. It plays very well and at a reasonable speed. Playing against XGR++ gives you a slight (very slight) improvement in playing strength but with a cost of much slower play.

If an interesting position arises you can always analyze it later by clicking on ++ or doing a rollout.

Thanks but that didn't answer either of my questions. As I pointed out I have a very fast computer. I always play against XG++ and it usually plays in a couple of seconds. I assume there is no way to get it to play stronger, I was just hoping. I would be happy if I could change the search interval to gigantic. The second question was what I should have led with as that was where I thought I could actually get help. I just don't understand how doing rollouts using variance reduction has value. Isn't that just asking the NN its opinion multiple times? How is that ever going to return an answer significantly different (and as good or better) than an XG++ analysis using gigantic?

You might be right and I mix up things, but IIRC I've read it or talked to a programmer about this in 2003 or 04. Your arguments are valid, but if you look at the situation at the start of the game, you have for the d1 queen about 20 fields to go to, and then you shoot the array to about the same amount of fields and you have 4 queens, so at the start you have a branching factor of 4*20*20 or about 1600 , 4 times the factor of GO! I don't follow amazons tourney, I don't know how they progressed and I assume that NN player could be a reasonable approach.
But back at that time, were the top chess programs did run on 4 single core CPU computers at the computer chess championship the branching factor was a challenge for the available hardware

The reason behind the usefulness of VR reveals if you look at the statistics behind. When you want a certain degree of confidence in your results you need a certain amount of games. The variance has the nasty property to enlarge that number, therefore it pays to do additional calculations to reduce the variance.
If you need 10 times the work to play a single game, but only need 1% of the games your 10 times as fast.

The move filter has only influences of the probability to overlook a good moves in the first place. If you evaluate all possible moves to 5 ply it will take a lot of time, even with XG, therefore bad moves are pruned quickly. Because the evaluation of the bots is quite well AND possible errors are mitigated by the dice BG bots can prune very aggressive.

Thanks again for the answer but again that didn't answer my question. I guess I didn't make myself clear. Here goes the long winded version: Isn't using variance reduction asking the NN to assign weights based upon ITS OPINION of the luck? I am not interested in doing a rollout just to get the same answer that I got from my normal analysis using XGroller++ gigantic more slowly. Speeding up a rollout to get the same wrong answer makes no sense to me. Presumably you are only doing rollouts when you suspect that the analysis returned was wrong. Why else would anyone do a rollout? If anyone has done a rollout using variance reduction and received an answer that differed by more than .01 from an XGroller++ gigantic analysis I would like too know the position so I could experiment for myself.