Tesla backgammon?

[Tim Chow]

I bought Robertie's new book and so got onto his mailing list.
One thing he mentioned recently is that the Tesla backgammon
program has apparently improved from being a terrible player
to being a challenging player. Does anyone here know anything
about this?

From a business point of view, it would seem much better to
have a weak program than a strong program, since you want to
make the customer feel smart.

---
Tim Chow

[bananab...@gmail.com]

I've never heard of it before right now but a quick Google search shows results as recently as the beginning of 2020. Reading the reviews it was hopeless and I mean hopeless. Random people reportedly beating it 13-0, 54-0, haven't lost a single game, sounded like it picked any random move. There's making the customer feel smart and there's making them feel like they're playing a worthless piece of software who doesn't know how to play the game.

On May 6th an update was released and it has been trained playing "20 million + games". Quick scan I don't see any reports of how strong it is but 20 million games, I'd still take Elon Musk's money if someone wanted to bet on it.

Stick

[bgbl...@googlemail.com]

20 Million games is enough for a quite decent player...... if you know what you do. Maybe they should ask one who knows what to do ;)
But as Tim pointed out: for the average customer to keep happy playing strength might not the right way to go. Does one remember the "Big Bang Board Games" Apple put on their Macs a decade or so ago. Much Bling-Bling and abysmal game play....

ciao
Frank

[Tim Chow]

On Saturday, May 30, 2020 at 9:20:32 AM UTC-4, bgbl...@googlemail.com wrote:
> 20 Million games is enough for a quite decent player......
> if you know what you do. Maybe they should ask one who knows what to do ;)

Yes, I was rather surprised when I first learned that 20 million games was
enough training data for a decent bot, because I was naively guessing that
it would require ten times as much as that. But apparently, even AlphaZero
chess used only 44 million chess games, so these neural nets are somehow
able to generalize incredibly well from (what seems to me like) relatively
little data.

> But as Tim pointed out: for the average customer to keep happy playing
> strength might not the right way to go. Does one remember the "Big Bang
> Board Games" Apple put on their Macs a decade or so ago. Much Bling-Bling
> and abysmal game play....

I have posted here before about the terrible level of play of the backgammon
game on JetBlue. But Stick has a point...if the computer player is *too*
weak then that can also be unappealing.

---
Tim Chow

[Tim Chow]

On Saturday, May 30, 2020 at 12:17:25 PM UTC-4, I wrote:
> Yes, I was rather surprised when I first learned that 20 million games was
> enough training data for a decent bot

Just checked Tesauro's paper and found that TD-Gammon Version 2.1 used only
1.5 million games. That wouldn't be considered a strong bot today but it
still illustrates the point that 20 million games is a reasonable amount of
data.

---
Tim Chow

[Bradley K. Sherman]

From scratch, or after having compiled a database of perfect
play in noncontact positions?

-bks

[Tim Chow]

On Saturday, May 30, 2020 at 1:05:43 PM UTC-4, Bradley K. Sherman wrote:

TD-Gammon at least didn't rely on a precomputed database of noncontact
positions. I don't think that most bots use such a database to train
the net, and my intuition is that it wouldn't help much.

But when you say "from scratch," there's some question as to what that
means. Versions of TD-Gammon after the first did make use of hand-crafted
features designed by human experts.

---
Tim Chow

[bgbl...@googlemail.com]

Am Samstag, 30. Mai 2020 22:48:28 UTC+2 schrieb Tim Chow:
> But when you say "from scratch," there's some question as to what that
> means. Versions of TD-Gammon after the first did make use of hand-crafted
> features designed by human experts.

Yes, it's not totally self learned, but the "expert inputs" are nothing spectacular. At least for GnuBG, BGBlitz and TD-Gammon it's more or less the stuff Berliner has published some decades ago :)

>That wouldn't be considered a strong bot today

I guess that depends what you regard as strong. I'm pretty sure it's PR is around or below 2 and the level of play comparable to Jellyfish. Given that, at least to my best knowledge, it belongs in the top ten. IMHO the order is as follows:

XG
GnuBG
BGBlitz
Snowie
Jellyfish / TD-Gammon

not too bad for a 30 year old SW ;)

BTW the training process for BG is pretty fast because it's ideally suited for the usual architecture (MLP with 1 hidden layer) and even with some ten thousand games the play is already more than reasonable.
For levels below "expert" I use a pretty small net with only 45000 games training and even then I add random noise to avoid casual player frustration.
The real AI is trained between 75 and 100 million games, but the increase in strength is really small if you are beyond some millions.

[WF Frantz]

Yes. The game maker claims that he used AI to improve, but that is total bull. It now cheats like crazy.

Even the 2023 version is rigged, but not how you think. After nearly 10,000 dice rolls, this is what I found when keeping track of rolls and getting back onto the board after being captured (clearing the bar):

The designer claims that the new game was put into learning mode and got a lot better, which is a total lie. He never did that. It got better by improving how it hides cheating.

Doubles are now OK. But ...

Getting out of the Bar (pieces taken) is rigged in favor of the computer. The computer beats the expected odds of getting back onto the board while the player is more likely than expected to not get back onto the board.

It is more subtle about the dice. It equally rolls more 4s for the computer and more 1s for the player. In Super Tesla mode, the dice rolls are now about 13% higher for the computer, so the computer has higher odds of completing first. This is subtle. The average roll should be 7. But the computer average is 7.2 and player 6.8. And by increasing the number of 4s for the computer and 1s for the player, the 1s are a definite handicap because blockades are normally face to face making the use of 1s limited.

The game is still stupid on the end game, which is the only reason why players continue to beat the computer in Super Tesla mode about half the time. The game has little situational awareness as to: (a) minimizing a 2 or 3 point penalty for losing, (b) focusing on getting pieces home once all pieces have passed each other, and (c) prioritizing clearing home and balancing home when a piece can't be cleared.

I wish that Tesla would not reward game makers that use cheating especially when they claim that they used AI. I wish Tesla would reward game makers with a work ethic that represents the work ethic of Tesla.

[peps...@gmail.com]

FWIW, my intuition was also that the low tens of millions would be far too few for
complex games like chess or backgammon.

You can take the man out of combinatorial game theory, but you can't take combinatorial
game theory out of the man.

Paul

[peps...@gmail.com]

If I remember rightly, it was an open question whether TD-Gammon was better/worse/equal
when compared to the world's strongest humans.
Of course, with the huge luck factor in backgammon, it can be very hard to ascertain these things.
I think that now, there's a strong consensus that both XG and Gnu are much stronger than any human,
and I'm sure there are plenty of other bots that surpass human play.

Paul

[Frank Berger]

peps...@gmail.com schrieb am Samstag, 3. Juni 2023 um 10:38:45 UTC+2:

> FWIW, my intuition was also that the low tens of millions would be far too few for
> complex games like chess or backgammon.

How many games a human needs ;) When you play with nets that are trained ony a couple of thousand games, you can watch how they graps more and more concepts.

[Frank Berger]

peps...@gmail.com schrieb am Samstag, 3. Juni 2023 um 10:42:28 UTC+2:

> I think that now, there's a strong consensus that both XG and Gnu are much stronger than any human,
> and I'm sure there are plenty of other bots that surpass human play.

How many are plenty?
As in an earlier post:
- XG
- GnuBG
- BGBlitz
- Snowie
are very very probably better than any human on average

TD-Gammon/Jellyfish/Motif? Are you sure that these are better than *any* human. If Mochy, ZZ and Ueda play together as a team against them, on whom you would bet? I'm *very* unsure

Do I miss some other?

[Frank Berger]

WF Frantz schrieb am Freitag, 2. Juni 2023 um 23:23:53 UTC+2:
would you mind to make your data available? cheating with the average would be a more smart way to do it

[peps...@gmail.com]

Motif was terrible last time I played it. It may have improved since then.
I don't know these issues particularly well. I suppose "sure" was probably the wrong word.
The team hypothesis does not address the question because "better than any human" means
"better than any individual human".
It's plausible (at least to me) that a team of the 20 best humans could play better than XG.

Paul

[peps...@gmail.com]

On Sunday, June 4, 2023 at 10:21:48 AM UTC+1, peps...@gmail.com wrote:
....

> Motif was terrible last time I played it. It may have improved since then.

...

I wanted to test the latest Motif but I was stuck on the Java requirements.
I got this message: "If this message appears, then your web browser does not run Java and you will not be able to play Motif backgammon."

I'm sure this has absolutely nothing to do with Motif because other java-requiring sites exhibit exactly the same issue.
I'd like to get this working on Google Chrome but I couldn't get it working on Microsoft Edge or Firefox either.
I've tried googling instruction sets without success.
Most of the instructions appear to assume menu functions which don't seem to exist.
I am using Windows 64 x64

Many thanks for your help.

I do have some partial successes as follows:
Successful installation indicated as below:
java -version
java version "1.8.0_371"
Java(TM) SE Runtime Environment (build 1.8.0_371-b11)
Java HotSpot(TM) Client VM (build 25.371-b11, mixed mode)

Paul

[Frank Berger]

peps schrieb am Sonntag, 4. Juni 2023 um 12:55:44 UTC+2:
> I wanted to test the latest Motif but I was stuck on the Java requirements.
> I got this message: "If this message appears, then your web browser does not run Java and you will not be able to play Motif backgammon."

Java in the browser isn't supported in Browsers nowadays (for good reasons but this is a long long story)

I checked when Java was thrown out of java itself, but it was in Java 9 so your Java 1.8 should work. The program would be the "appletviewer"... but it don't
so before investing a lot of time I thought: let's try it with an old environment, but on my recent boxes no browser supports the Java-plugin and on the older (OS/2 in a VM, iMac from 2000) they couldn't handle the new algos for https/ssl :(

finally I checked it with curl whether the files are still there.... and they are. I downloaded them and packed them together at.
https://bgblitz.com/download/misc/Motif.zip

If you
- download it,
- unpack it at any convenient place
- cd the commandline in this directory
- and call
appletviewer ./motif.html

it should work (at least on a Linux and a Mac it works. To lazy to start WIndows).

BTW there is a webpage , where Hans-Jürgen Schäfer collects elo numbers from every BG bot he can get:
https://mustrum.de/blot.html

It is in German, but if you scroll down you'll find the elo numbers. I don't know his exact procedure, but the number looks plausible to me. The most bias is probably due to setting max elo to 2050.
Astonishing/amazing how bad some bots are....

best
Frank

[peps...@gmail.com]

Thanks a lot for your help. I'll let you know how I get on.

Paul

[MK]

On June 3, 2023 at 2:42:28 AM UTC-6, peps...@gmail.com wrote:

> On May 30, 2020 at 5:20:43 PM UTC+1, Tim Chow wrote:

>> On May 30, 2020 at 12:17:25 PM UTC-4, I wrote:

>>> Yes, I was rather surprised when I first learned
>>> that 20 million games was enough training data
>>> for a decent bot

What one deems "decent" is relative and also would
depend on being a decent bot at what kind of games.
Those were all cubeless one-pointers.

>> Just checked Tesauro's paper and found that
>> TD-Gammon Version 2.1 used only 1.5 million
>> games. That wouldn't be considered a strong
>> bot today but it still illustrates the point that 20
>> million games is a reasonable amount of data.

It wasn't a strong bot back then either. Not 1.5 nor
even 20 million "cubeless one-pointer games" can
be a "reasonable amount of data" for for cubeful
money and/or match play, unless some jackoffski
formulas, MET's, etc. are substituted for the lack
of training and unless "the small incestious circle
of gambling gamblegammon players" subscribed
to it.

> If I remember rightly, it was an open question
> whether TD-Gammon was better/worse/equal
> when compared to the world's strongest humans.

It was only compared to a few humans considered
strongest among a handful of gamblers who rated
themselves against one another but at least a real
comparison was done by those humans actually
playing against the TD-Gammon.

> I think that now, there's a strong consensus that
> both XG and Gnu are much stronger than any human

Consensus doesn't mean based on factual data. The
question is still open and even more so now than back
then because unlike with TD-Gammon, there has been
no serious comparing to humans through actual play
against those bots with incentives enough for humans
to really try to beat them.

Just the consensus among billions of insect brained
people that gods, angels, spirits, etc. exist doesn't
make their fantasies into facts...

MK

[MK]

On June 3, 2023 at 2:35:32 PM UTC-6, Frank Berger wrote:

> How many are plenty?
> As in an earlier post:
> - XG
> - GnuBG
> - BGBlitz
> - Snowie

Why do you keep injecting your garbage
bot BG-bzzt into the list at evert occasion?

Even though it's dated by now, Snowie still
belongs there because it's the only bot that
disagrees with Ex-Gee and Noo-BG at times.

Nobody other than yourself does hardly ever
mention, let alone praise, your garbage bot
BG-bzzt. :(

Add some unique features to it to make it a
little more useful and valuable, such as the
random rollouts that I've been suggesting for
some years now.

In BGO, they are talking about PC's that can
do 46 million positions/second in Ex-Gee.

The 1,296,000 trials random rollout that I did
on my 3.8 million positions/second PC had
taken less than 5 hours. That means with a
Ryzen 9 7950X CPU you can do a ten times
longer random rollout, i.e. 12,960,000 trials,
in just 4 hours.

If a 12 million 960 thousand trials of random
cubeful rollout cames up with different best
plays than a 1,296 trials Ex-Gee rollout with
4-ply checker and XGroller+ cube settings,
will you all not accept it as more credible..??

What are you all afraid of to not add a simple
feature to your bots that could demonstrate
them to be pieces of shit in a matter of a few
dozens of random play rollouts that would at
most take a few days...?

You already know it! You can delay accepting
it openly for now but you can't prevent it forever.
It's only a matter of time...

MK

[Frank Berger]

MK schrieb am Montag, 5. Juni 2023 um 11:29:54 UTC+2:

> Nobody other than yourself does hardly ever
> mention, let alone praise, your garbage bot
> BG-bzzt. :(

Not that I believe that you were reachable for arguments or facts, but feel free to make some matches with BGBlitz (I suggest 4- or 5-ply) or BGBlitz against itself and feed them to XG. You don't regard PR as reasonable IIRC but Snowie is a bit over 1 and BGBlitz at 5-ply is around 0.6 and Mochy is about 2.3.

Or you may remember the Othello 2023 Quiz posted a few days ago? What you call "garbage" solves
- 8 out of 10 with 1-ply
- 9 out of 10 wth 2-ply
- 10 out of 10 with 5ply (the 1st position is really a beauty)
I'm confident that you wont regard that as showing anything, but others might find that interesting. BTW one doesn't need a license to check that, just drag&drop the XGID on the board.

[MK]

On June 5, 2023 at 4:49:53 AM UTC-6, Frank Berger wrote:

> MK schrieb am Montag, 5. Juni 2023 um 11:29:54 UTC+2:

>> Nobody other than yourself does hardly ever
>> mention, let alone praise, your garbage bot
>> BG-bzzt. :(

> ... feel free to make some matches with BGBlitz

> (I suggest 4- or 5-ply) or BGBlitz against itself and
> feed them to XG. You don't regard PR as reasonable
> IIRC but Snowie is a bit over 1 and BGBlitz at 5-ply
> is around 0.6 and Mochy is about 2.3.

My point is that your bot isn't popular. I don't care
if it's almost as strong as Ex-Gee. Even yourself
measure your bot against Ex-Gee, Noo-BG, Snowie.
They're all the same. Your bot doesn't offer anything
significantly different than other bots.

You seem to be making quite an effort to improve
and keep your bot current but you won't even take
a small suggestion that could help it break away
from the inbread dog pack. :( At time, I try to help
you out with some ideas to no avail...

> Or you may remember the Othello 2023 Quiz posted
> a few days ago? What you call "garbage" solves
> - 8 out of 10 with 1-ply
> - 9 out of 10 wth 2-ply
> - 10 out of 10 with 5ply (the 1st position is really a beauty)

Big fucking deal. They were all rolled out with Ex-Gee
"1296 Games rolled with Variance Reduction.
"Moves: 3-ply, cube decisions: XG Roller

As if I had nothing better to do, I rolled them all using:
"1296 Games rolled with Variance Reduction.
"Moves and cube decisions: 1-ply

Guess what? Ex-Gee 1-ply got 10 out of 10 right.

(BTW: this validates Tim's claims that Othello pics
positions that won't vary based on bot and/or rollot
settings).

Furthermore, if the quiz was done using BG-bzzt, it's
very likely that your 1-ply would get them all right.

You people just can't understand the absurdity of it
all... :(

But more importantly, why are you ignoring/avoiding
my suggestion to you?

I just looked at your bot's rollout settings. I only see
1, 2, 3-ply, no 5-ply. But that's not important. Random
rollouts can make your bot stand out from the rest.
How can't you see that? And when it's so easy to do.

All you need to do is modify one line and add 5 lines
to your code. Here are the steps:

1- Modify the line that sets the size of your drop down
list array from 3 to 4.

2- Add a new line to add a new rollout level, (i.e. random),
to the array.

3- I assume you have a block of code, perhaps with calls
to subroutines, which decides if the player can/should
initiate a cube action and if/how he must respond to a
cube action. Just before that block add a line to execute
it contitionally, (i.e. IF ROLLOUT-TYPE NOT = "RANDOM")
and right after it add a line to make a random cube
decision, (i.e. ELSE CUBE-ACTION = RAND(0 or 1)). That's
it. Just let the rest of the cube logic execute as normal.

4- To decide how to move, I assume your code first fills
an array with all possible moves and then evaluates and
ranks them to pick the best move. Right after filling the
array, again add a line to execute that block of code
conditionally, (i.e. IF ROLLOUT-TYPE NOT = "RANDOM")
and just before the code that executes the chosen play,
add a line to pick a random move from the array, (i.e.
ELSE BEST-MOVE = ARRAY-ELEMENT(RAND(1 to ARRAY
SIZE). That's it. Just let the rest of the checker play logic
execute as normal.

Okay, my sample lines to add are very sloppy but I'm
sure you get the idea. Just modify 1 line and add 5
new lines. How long would this take for you to do?
20 minutes? 30 minutes? 1 hour?

What's your excuse for refusing to add this simple
yet potentially jackoffski cube skill debunking,
gamblegammon shaking feature to your garbage
bot...??

Ex-Gee is obviously an abandoned product by now
but I don't understand why Noo-BG team doesn't add
this feature to their bot, which would be just as easy
to do as adding for you to add it to your bot.

Remember that you all won't be able to prevent the
unevitable. Save yourselves the agony and the pain.
Bite the bullet. Get done and get over with it. Sooner
the better for you all. Trust my friendly advice... ;)

MK