Discussion:
Tesla backgammon?
(too old to reply)
Tim Chow
2020-05-23 16:17:15 UTC
Permalink
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
b***@gmail.com
2020-05-23 16:55:20 UTC
Permalink
Post by 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
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
b***@googlemail.com
2020-05-30 13:20:30 UTC
Permalink
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
2020-05-30 16:17:24 UTC
Permalink
Post by b***@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 ;)
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.
Post by b***@googlemail.com
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
2020-05-30 16:20:42 UTC
Permalink
Post by Tim Chow
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
2020-05-30 17:05:42 UTC
Permalink
Post by Tim Chow
Post by Tim Chow
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.
From scratch, or after having compiled a database of perfect
play in noncontact positions?

-bks
Tim Chow
2020-05-30 20:43:22 UTC
Permalink
Post by Bradley K. Sherman
Post by Tim Chow
Post by Tim Chow
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.
From scratch, or after having compiled a database of perfect
play in noncontact positions?
There are others who answer your question more definitively than I can,
but I'd be surprised if that made that much difference. Let's suppose
that a neural net's noncontact checker play is poor. How much equity
will that cost it in the long run? Some, of course, but I would guess
not a whole lot, as long as it's not doing things like refusing to bear
off checkers when it's obviously correct to do so.

Noncontact cube play might matter more, but I'd think that Janowski would
still do a good job even if you didn't have a database telling you how to
handle the cube.

---
Tim Chow
Tim Chow
2020-05-30 20:48:27 UTC
Permalink
Post by Bradley K. Sherman
Post by Tim Chow
Post by Tim Chow
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.
From scratch, or after having compiled a database of perfect
play in noncontact positions?
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
b***@googlemail.com
2020-06-04 22:20:23 UTC
Permalink
Post by 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 :)
Post by Tim Chow
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.
Loading...