Incredible miss and backgames against bots

[PersianLord]

This position was taken from a match between a fibster (White) and MonteCarlo bot (gray)  this afternoon.



The human player (White) had 97% chance to hit: 35 of all rolls other than 1-1, and guess what happened, he rolled 1-1.  .  Every 2,3,4,5 and 6 could have done the job for White, who apparantly had played a very excellent and well-timed backgame, yet failed to hit.

I just thought that this position might be interesting to post. After all, it teaches us that NOTHING is guaranteed in backgammon. Unfortunately, I was banned by the white player and thus was not able to watch the rest of match, but I guess he should have won the match, considering his excellent position (spare checker on 5-pt anchor, strong board, 3-anchors and terrible timing of the damn bot.)

By the way, haven't you noticed that almost all of the bots seem to have problems in facing backgames?

[spielberg]

All the bots will have (very small) problems against backgames as they're a relatively small portion of the universe of matches and thus they'll be less well educated in that subset. The magnitude of the bots errors will be nowhere near that which Hardy once identified when he claimed (almost certainly correctly given what I know of him) that he'd win 100% of the time against GnuBg when both players start with 15 pieces on the bar but will arise from a similar "blindspot". Whether the errors are abitrageable is an extremely difficult question - I suspect not yet know NIHI (who's excellent generally (please note that's solely in term's of his play) and particularly at backgames) tries.

[adrian]

Issuing a look NIHolymypic command every 2 seconds will allow you to manually compile a book about NIHI's matches, despite of the gag and blind. Publish it , it may be interesting!

[PersianLord]

All the bots will have (very small) problems against backgames as they're a relatively small portion of the universe of matches and thus they'll be less well educated in that subset. The magnitude of the bots errors will be nowhere near that which Hardy once identified when he claimed (almost certainly correctly given what I know of him) that he'd win 100% of the time against GnuBg when both players start with 15 pieces on the bar but will arise from a similar "blindspot". Whether the errors are abitrageable is an extremely difficult question - I suspect not yet know NIHI (who's excellent generally (please note that's solely in term's of his play) and particularly at backgames) tries.

Good point, Steve.

I'm not good at backgames and hate to play them, but believe me, whenever I tried, I did better against GNUBG. You can test it steve, start a well-planned backgame and you'll see GNUBG makes some strange blunders. It seems that it doesn't consider timing and re-circulation when defending against a backgame. It hits as many blots as it can, even when you've made 2 points in his board. Thus it usually endes with many checkers out of play and a terrible timing, much like the position I posted.

I can bet the White has won that match, other than a drop out of rage!

[PersianLord]

Issuing a look NIHolymypic command every 2 seconds will allow you to manually compile a book about NIHI's matches, despite of the gag and blind. Publish it , it may be interesting!

Thanks my dear friend, but I'm not really interested in such a book, cuz I know already how to defeat sub-1500 opponents  .

The cause for posting this position was that the white player (whom I didn't like to write his FIBS nickname because of my deep respect for privacy) shouted and asked us all to check his position. I got interested and took a look. Then thought that it might be nice for other fellow fibsters to take a look as well and here we are!

Regards

[Noah]

It's like throwing a 21 when only a pip from 4 is needed to WIN.

Further nothing new.... Drama of the backgame should be incalculated.

[dorbel]

It is an opinion often heard that "Bots don't know how to play backgames" and by implication don't know how to defend against them either. Is there any actual evidence of this? Snowie, Gnu and bgBlitz all seem to play them pretty well to me.

[PersianLord]

It is an opinion often heard that "Bots don't know how to play backgames" and by implication don't know how to defend against them either. Is there any actual evidence of this? Snowie, Gnu and bgBlitz all seem to play them pretty well to me.

There is an article that claims that snowie4 is better in defending against backgames:

http://bgsnowie.com/backgammon/articles.dhtml?id=56

[dorbel]

That article compares Snowie 4 to Snowie 3. What I really want to know is, why do we so often hear that the claim that bots dont understand backgames? Where is the evidence?

[blitzxz]

That article compares Snowie 4 to Snowie 3. What I really want to know is, why do we so often hear that the claim that bots dont understand backgames? Where is the evidence?

This is a big problem... How can you prove that bot is playing badly? Maybe rollout. But that might be wrong as well if there is systematic errors. You might believe some (well known) expert player but that's far from evidence to me.

I think bots play normal backgames well and I have learnt a lot about backgames from gnu. I think that gnu is right when it often hits checkers. If opponent already has a good timing you might as well send more checkers to bar because that increases gammon & backgammon chances. But gnu doesn't hit if opponent has bad timing. And gnu tries to ruin the timing by slotting to home board.

The only bad thing which I have found out is that when gnu is in some sort of super backgame (3-points and more) it will slot like crazy. Gnu doesn't care how many checkers are send behind opponent bar. In one of those rare games gnu had almost all his checkers in my home board. I did also long rollouts (with gnu..) and they showed that gnu's recycling strategy was not correct.

And I don't know about snowie since I don't have it.

[PersianLord]

That article compares Snowie 4 to Snowie 3. What I really want to know is, why do we so often hear that the claim that bots dont understand backgames? Where is the evidence?

The article clearly shows that bots have been in trouble facing backgames for a long time:

"Overall Snowie 4 proves to be much better in backgame positions, in respect to the checker play and cube actions. Ambitious Backgammon players can learn a lot from Snowie in this domain. It is the first time, that a computer program handles backgames well!"

For example here, Snowie 3 will play a foolish 24/20*/14:



But Snowie 4 will correctly play 24/14 without hitting. When the 3rd version of Snowie was so bad handling the backgames, it's quite reasonable to think that how terrible might have been the earlier versions. So we can claim that it's true that bots have, or stil had, troubles in handling backgames. And still I believe that they can't handle them as good as they perform other sterategies.

Regards

[dorbel]

Yes I agree that traditionally bots were less efficient in backgames than in other types of games, but so are 99% of humans! My point is that many people seem to think that if they can get the bot into a backgame, then they will be able to outplay it, but I don't believe that, nor have I ever seen any evidence to support the theory.
I accept, because Levermann says that he has rolled it out, that not hitting is better in the position that he shows, but i wouldn't say that it is obvious over the board.
Anybody got a position where a modern bot makes a blunder (a move that costs 0.11 in money equity)?
Blitzxz correctly identifies a problem with rollouts, which is that you can't trust the bot to do a rollout if it is making a fundamental error to begin with. Thus if you want to compare a hitting play with a non hit in a position where the evaluation prefers the non hit, the rollout is of dubious value. It will play the non hit variation consistently and of course it will hit on the first roll if you ask it to do so, but it won't hit later. You see the problem.

[lewscannon]

I can't resist this. What no one has pointed out is the infernal 'luck' that these damn bots seem to have. I can't count the number of times when I can see an impending disaster coming 3 or 4 rolls away in a match against a bot, and there's absolutely nothing I can do to avoid it, as the dice railroad me with highly improbable sequences, whereas it just doesn't seem to happen nearly as often against regular players. I know this statement will have it's detractors (hi, dorbel! ), but after all my Fibs experience, it just seems to happen way more against bots than humans.

Just because you're paranoid doesn't mean that everyone isn't out to get you.

We now return you to your regularly scheduled program.

lews

[PersianLord]

I can't resist this. What no one has pointed out is the infernal 'luck' that these damn bots seem to have. I can't count the number of times when I can see an impending disaster coming 3 or 4 rolls away in a match against a bot, and there's absolutely nothing I can do to avoid it, as the dice railroad me with highly improbable sequences, whereas it just doesn't seem to happen nearly as often against regular players. I know this statement will have it's detractors (hi, dorbel! ), but after all my Fibs experience, it just seems to happen way more against bots than humans.

Just because you're paranoid doesn't mean that everyone isn't out to get you.

We now return you to your regularly scheduled program.

lews

Well done.

Actually my initial title for this topic was "incredible miss and an old question". The old question was :" Do FIBS bots cheat?", lol.

I still can't realize why bots get their most desired rolls, most of the times, but let us imagine their dices are really random.

Regards

[Zorba]

Human beings tend to be very bad at dealing with probabilities and statistics. Especially dealing with very low or very high probabilities is difficult for us, at least at the psychological/intuitive level.

Rolling 1-1 here is not a miracle. It's just one of the 36 permutations of rolling two dice. It's almost a 3% chance to roll 1-1, every time you roll two dice, and a bit more than 97% to not roll it. We tend to translate that to "can't miss" and "will hit", because it's hard to deal intuitively with the unlikely 3% scenario. But it should happen once every 36 times, and given the amouhnt of dice most players roll on FIBS (especially combined together!), things like this happen all the time. So this might feel like "an incredible miss", but it really isn't. It's just a case of rolling the unluckiest roll, a 1/36 chance. It will happen at least once in quite a lot of your backgammon games. This particular position just happens to have one very bad roll, something which is quite common for backgames (for both sides even).

Another way of looking at it: if you had rolled something else and hit, would you consider yourself lucky that you didn't roll the dreaded 1-1? Or would you have found that "normal" or not even have realised there was a chance to miss?

Here's a quick rollout of the Snowie 4 position (hit or not?), GNUBG 0-ply cubeless:


    1. Rollout          24/14                        Eq.:  +0,432
        62,5%  21,8%   0,9% -  37,5%   4,5%   0,1% CL  +0,432
      [  0,2%   0,2%   0,1% -   0,2%   0,1%   0,0% CL   0,006]
        Full cubeless rollout (trunc. at one-sided bearoff) with var.redn.
        1296 games, Mersenne Twister dice gen. with seed 825414521 and quasi-random dice
        Play: 0-ply cubeless prune
    2. Rollout          24/18 13/9                   Eq.:  +0,379 ( -0,053)
        59,5%  24,2%   1,4% -  40,5%   6,3%   0,3% CL  +0,379
      [  0,3%   0,2%   0,1% -   0,3%   0,2%   0,1% CL   0,008]
        Full cubeless rollout (trunc. at one-sided bearoff) with var.redn.
        1296 games, Mersenne Twister dice gen. with seed 825414521 and quasi-random dice
        Play: 0-ply cubeless prune
    3. Rollout          24/20* 13/7                  Eq.:  +0,343 ( -0,089)
        53,6%  30,4%   2,4% -  46,4%   5,4%   0,2% CL  +0,343
      [  0,3%   0,2%   0,1% -   0,3%   0,2%   0,1% CL   0,008]
        Full cubeless rollout (trunc. at one-sided bearoff) with var.redn.
        1296 games, Mersenne Twister dice gen. with seed 825414521 and quasi-random dice
        Play: 0-ply cubeless prune
    4. Rollout          24/20*/14                    Eq.:  +0,337 ( -0,095)
        53,9%  29,4%   2,2% -  46,1%   5,5%   0,2% CL  +0,337
      [  0,3%   0,2%   0,1% -   0,3%   0,2%   0,0% CL   0,008]
        Full cubeless rollout (trunc. at one-sided bearoff) with var.redn.
        1296 games, Mersenne Twister dice gen. with seed 825414521 and quasi-random dice
        Play: 0-ply cubeless prune


So, hitting is indeed quite bad, costing you lots of wins. However, it does win more gammons (and some backgammons); just not enough to make it a good play.

N.B. I think this rollout is good enough for ballpark figures here (i.e. hit or not), but if you want more accurate numbers you really should use 2-ply checker play.

[PersianLord]

Another way of looking at it: if you had rolled something else and hit, would you consider yourself lucky that you didn't roll the dreaded 1-1? Or would you have found that "normal" or not even have realised there was a chance to miss?

Not really. When something is likely to happen 97% of the times, it's occurance would not be considered lucky. That would be just normal.

So, hitting is indeed quite bad, costing you lots of wins. However, it does win more gammons (and some backgammons); just not enough to make it a good play.

I had pointed to this earlier, that even without hitting this roll, white is still the favorite to win the match. But my main focus was actually on this question: "why bots, here MonteCarlo, are not good at backgames?"

Regards

[socksey]

"why bots, here MonteCarlo, are not good at backgames?"

Because they have to give us some miniscule chance of beating them!   

socksey



"Don't tell people how to do things. Tell them what to do and let them surprise you with their results." - George S. Patton

[blitzxz]

I can't resist this. What no one has pointed out is the infernal 'luck' that these damn bots seem to have. I can't count the number of times when I can see an impending disaster coming 3 or 4 rolls away in a match against a bot, and there's absolutely nothing I can do to avoid it, as the dice railroad me with highly improbable sequences, whereas it just doesn't seem to happen nearly as often against regular players. I know this statement will have it's detractors (hi, dorbel! ), but after all my Fibs experience, it just seems to happen way more against bots than humans.

The reason is that bots play better. Humans only focus on their opponents good luck not realizing their own luck which is called skill. Bots maximaze their chances win so they seem luckier as you lose more. Selective observation. That's like those slot machine players who always remember when they have won but never remember their loses. I complitely agree with Zorba. The whole gambling industry is based on the fact that humans can't figure out propabilities. If every one could no one would play a game for money where they are expected to lose. Sometimes you lose, sometimes you win, but can you figure out the big math: how are you doing in long run? Easy to say but millions of gamblers play by their own logic. Who has ever played lottery?  Not rolling the 1-1. Yes that's lucky. It's over the expected value. 

[Yvon]

This position was taken from a match between a fibster (White) and MonteCarlo bot (gray)  this afternoon.



The human player (White) had 97% chance to hit: 35 of all rolls other than 1-1, and guess what happened, he rolled 1-1.  .  Every 2,3,4,5 and 6 could have done the job for White, who apparantly had played a very excellent and well-timed backgame, yet failed to hit.

I just thought that this position might be interesting to post. After all, it teaches us that NOTHING is guaranteed in backgammon. Unfortunately, I was banned by the white player and thus was not able to watch the rest of match, but I guess he should have won the match, considering his excellent position (spare checker on 5-pt anchor, strong board, 3-anchors and terrible timing of the damn bot.)

By the way, haven't you noticed that almost all of the bots seem to have problems in facing backgames?

Hi this is my first post here. I agree with the previous posters (Zorba et al) that the 1-1 roll was just bad luck, it proves nothing really.

HOWEVER this is not what usually happens at Fibs. What happens is a SERIES of bad rolls for one player combined with either neutral or a series of best rolls for the other. This causes unbelievably high impropabilities that cannot be attributed to random dice. ****

OK in the above example we had the one and only bad roll 11 i.e with a probability 1:36.
Let's suppose the black player then rolls one of his best rolls 12,14,16, 21,23,24,25,32,34,36,41,43,54,56,61,63,65. The chance for this is 17:36 almost 50%.Not bad at all. Then white rolls one of the worst again 55, or 66. This is a chance of 2:36.

We have already reached a propability of this happening of 1/36*17/36*2/36 of 34:46656 or 1:1372
I won't be surprised if black then rolls some neutral rolls and then when ready, a very usefull 66 while white either rolls neutral or more catastrophic doubles.

*** So if this cannot be atrributed to random dice on what the hell can be atrributed? In my oppinion all games at FIBS are monitored by a computer program which at some stage (usually at critical positions) interferes and decides who the winner will be. Don't you dare double in such situations, you will end up crying.

I am going to open up a forum for discussing specifically this issue. I don't expect Fibs to ever admit this  happening, so my only hope is for people to submit their files, the number of games they played, and see if such high improbabilities fall within range.

I personally have played about 1000 games there, i have lost more than a 200 under such bizare conditions.

NB. Has anybody noticed how frequent is your rolling 66 at the BEGINING of the game (say within the first 7 rolls) when they HIT you? Is the chance for this happening a plain 1:36??? No it is not! Look at the conditions (within the first 7 rolls + hit you)

[Yvon]

Now if I may add my 2 cents on the matter of the bots been lucky (or rolling themselves a series of best rolls) or not, I would definetely say they are not. There are 6(?) bots at Fibs I think, 3 of them are just average players. Strangely I have never seen a series of high improbability best rolls from a bot, while i have seen too many while playing with humans. Don't know why,propably the Fibs server doesn't mess around when bots are playing 
Just lets them do their own messing which luckily is not as much as one would expect from Fibs server 

Gnu imo is not such a good player I don't know why so many people rely on its analyses and stuff. How reliable could they be??