AlphaZero

[Tim Chow]

DeepMind recently announced that its AlphaZero program was able to
produce superhuman performance not only for go (that was older news)
but also chess and shogi. This is a remarkable result given that
neural nets and Monte Carlo tree search had been tried before for
chess (though possibly not in concert) but without being able to
beat alpha-beta search.

Of course, neural nets "conquered" backgammon a long time ago.
However, superbackgames (along with complicated containment play
and prime-rolling all the way around the board) have remained a
sticking point for current bots such as XG and GNU. It would be
interesting to see if AlphaZero could overcome this particular
limitation, and if so, what effect that would have on its evaluation
of more "normal" backgammon play.

There is another interesting avenue that has been relatively
unexplored to date by the AI community, which is the exploitation
of weaknesses in the opponent's play. It's unclear to me whether
AlphaZero could be adapted to attack this problem. An obvious
difficulty is that simulating a particular opponent (say, for the
sake of argument, Stockfish, a leading chess engine) takes too much
computer time. One could of course pre-compute a corpus of
Stockfish games and give it to AlphaZero to "study," but a static
corpus is necessarily rather small and there is always the danger
of overfitting. Still, one virtue of neural nets is that they
are supposed to be able to generalize well from limited data, so
maybe the problem is not intractable.

---
Tim Chow

[bananab...@gmail.com]

Didn't Xavier state at one point who knows how long ago that he could create a bot that was (much) stronger at those areas you're speaking of but it would be at the cost of the normal engine? I seem to remember something along those lines but it doesn't really interest me so I didn't retain much.

Stick

[Michael]

On Monday, December 11, 2017 at 8:40:03 AM UTC+2, bananab...@gmail.com wrote:
>
>
> Didn't Xavier state at one point who knows how long ago that he could create a bot that was (much) stronger at those areas you're speaking of but it would be at the cost of the normal engine? I seem to remember something along those lines but it doesn't really interest me so I didn't retain much.
>
> Stick

What surprised me is the high expectation of people at BGO that AplhaZero (if modified to play BG) would actually beat XG.
If this ever happens I wouldn't be surprised seeing people throwing away XG and spending $200-400 buying the new bot!

And then what about all those people who wrote books-done rollouts with XG to prove something? What if A0 proves XG wrong on a significant %?
Get ready for another 20 year Murat session :-)

[bananab...@gmail.com]

The last thing I'm worried about is a new bot coming along proving XG wrong on a significant % of plays. That's not going to happen. XG, GNU, Snowie ... all stronger than the best human players under normal conditions. The next bot will be at most like all the former have been, a slight improvement over its predecessor.

Stick

[Tim Chow]

On Monday, December 11, 2017 at 1:40:03 AM UTC-5, bananab...@gmail.com wrote:
> Didn't Xavier state at one point who knows how long ago that he could create
> a bot that was (much) stronger at those areas you're speaking of but it would
> be at the cost of the normal engine? I seem to remember something along those
> lines but it doesn't really interest me so I didn't retain much.

Yes, this is true. But since we seem to be far from fully understanding such
positions, it's hard to know how much improvement is possible. Nobody knows
how to play close to optimally when you have, say, 9 checkers back and 6 other
checkers scattered around the board trying to contain two of the opponent's
stragglers. People have a strong tendency to overestimate how close they are
to "solving" a strategy game. I've seen this in go, chess, Othello, checkers,
and backgammon---people confidently pronounce that current understanding of
the game is nearly perfect, only to be proved spectacularly wrong.

> The next bot will be at most like all the former have been, a slight
> improvement over its predecessor.

I think that this will be true for most "normal" positions. But for
superbackgames, the improvement could be enormous. If that is true, then
there is the potential for significant improvements in less extreme backgames.
Of course, backgames don't come up that often even if you try to steer for
them. But it is conceivable that a bot of the future might be able to play
XG a lot, detect its weakness in superbackgames, and then adjust its play to
take advantage of that weakness, even in (say) 15-point matches. That would
be pretty cool to see.

Anyway, if someone wants to try to convince the DeepMind team to spend some
time on backgammon, the first thing to tell them is that current bots
absolutely suck at superbackgames. Being able to say that their program
crushes existing programs, even if it's a restricted type of position, is
the sort of thing they're aiming for.

---
Tim Chow

[BlueDice]

Oh no!
Another life term lol
--
BD

[Paul]

I thought that the computer revolution in chess largely vindicated human
play, rather than proving theory "spectacularly wrong." As far as I can
recall, the computer surprises were relatively minor. For example,
queen and king against rook and king is a highly difficult technical win,
rather than an easy win. It was wrongly thought to be an easy win because
good defence was so rare. All the major openings are still intact (I think)
although a few sub-sub-sub variations may be busted. The Queen's gambit,
semi-Slav, King's gambit, Ruy Lopez, Modern Defence, French Defence,
Caro-Kann etc. are all about as good as they were always thought to be.
It's not like backgammon where 8/2 6/2 in the opening went from being seen
as a "beginner's move" to optimal or near-optimal.

Paul

[Tim Chow]

On Monday, December 18, 2017 at 4:19:25 PM UTC-5, Paul wrote:
> I thought that the computer revolution in chess largely vindicated human
> play, rather than proving theory "spectacularly wrong."

What I said was "spectacularly wrong" was the belief that current
understanding is nearly perfect.

This sort of thing preceded computers. Capablanca thought that chess
would be played out within a few decades and that games would all end
in draws.

---
Tim Chow

[mu...@compuplus.net]

On December 11, 2017 at 3:09:48 AM UTC-7, Michael wrote:

> What surprised me is the high expectation of people at
> BGO that AplhaZero (if modified to play BG) would
> actually beat XG.

From my angle, when I read this, I say "So what? Even
I can beat XG"... :)

From your angle, I would bet all I have that an AI or
even a real NeuralNet bg bot would decimate XG.

It's laughable that some people try to elevate XG to
the level of AplhaZero.

XG is not only light years away from being AI but the
only NeuralNet that there is in it is what it inherited
from the very first training of TD Gammon.

After that came all the arbitrary bullshit that got
hardcoded into Jellyfish, Snowie, Gnubg and XG, which
now make the bulk of bg bots, with a tinge of NeuralNet.

> If this ever happens I wouldn't be surprised seeing
> people throwing away XG and spending $200-400 buying
> the new bot!

Then the "XG team" (Xaviar, Cyborg, Kazoo, Suchi, etc.?)
would steal it and resell it for $50. So, don't none of
you worry... ;)

> And then what about all those people who wrote books-done
> rollouts with XG to prove something? What if A0 proves XG
> wrong on a significant %?

These are not questions of "what if" but "when"...!

Unfortunately, or fortunately, people like "AplhaZero team"
won't bother with bg which is by now nothing but a gambling
tool for a handful of mentally ill gambling scumbags (a few
thousand word wide??)

Those miserables, ranking from midget to giant :)", are so
fucking desperate for attention that everytime they see bg
mentioned in a magazine article, a movie, etc. that share
and celebrate...! :))

I guess I was over reacting to their goals/efforts by trying
to warn people to stay away from the circle of scumbags and
keep their kids away from them also. Apparently they can't
reach an audiance of more than a few anyway...

> Get ready for another 20 year Murat session :-)

Anything I can do to give you backstabbing cocksucker and
your ilks a reason to exist... ;)

MK

[Paul]

On Wednesday, February 7, 2018 at 5:16:10 AM UTC, mu...@compuplus.net wrote:
...

>
> Unfortunately, or fortunately, people like "AplhaZero team"

> won't bother with bg ...

This part is actually true. Lots of games are "bigger fish to fry"
from the perspective of the pioneers in game theory by neural networks.
They would be much more interested in chess/poker/shogi etc.

XG is already clearly the best player, and success is perceived about
becoming the best in as many areas as possible, not in improving in areas
where you already lead. Also, bg doesn't have the prestige of
many other games.

Murat is a master of deep analytical thinking, and has made enormous
contributions to backgammon theory.

However, the community lacks people with his expertise and genius and
that might be part of the reason bg lacks prestige. For those who want
to see bg theory further developed, a way to make progress might be
to examine what Murat does and to follow his example. What would Murat do?

Paul

[mu...@compuplus.net]

On February 7, 2018 at 4:24:54 AM UTC-7, Paul wrote:

> Murat is a master of deep analytical thinking, and
> has made enormous contributions to backgammon theory.

Thank you. I knew I would get recognition eventually.

MK