Now, for cube decisions, what if we don't use any
jackoffski formulas but simply make them crudely
on cubeless equities..?
If Deepmind will ever do something like DeepBG I'm very confident that they won't come up with such a success as in Go and chess.keynote how to parallize learning he said that they had other people that did this (so the 18 were only involved with Go). Compare this to leisure time programmers (BGBlitz, GnuBG) or a single full time programmer (Xavier?). The hardware that learned "
- chess and go are two dimensional games. The idea to use NN that are used in picture recognition was the key idea. BG instead is only a 1-dimenional game so the achievements of deep learning don't achieve much
- if you take into account what a amount of resources deepmind throws in, they would surely come out with something better, but not washing the floor with any of XG, BGBlitz or GnuBG. The alphago team was 18 people IIRC. When Aja Huang was asked on a
- deep mind always tries to do thing that haven't been done yet.Unless one of their Cxx get's mad on BG they wont do BG.
More in the AlphaZero spirit would be to train a separate net
for each match score. Start by training neural nets for DMP,
and 2-away Crawford. No cube here to worry about. You can
then figure out what the match-winning chances are at 2-away
Crawford, by doing Monte Carlo simulations.
The first score where you really need to consider the cube
would be 2-away post-Crawford. If you're willing to cut
corners then you could cheat and use human intervention
here, but it would be interesting to see if the neural net can
figure out the cube on its own. In particular, does it
play on for the gammon if it's the 2-away player and wins
the opening roll with 31 and the opponent rolls 62 or 63?
It would be interesting to see a neural net try to figure out
the cube strategy on its own here.
(Of course if you're willing
to cut corners, you could use the "standard" match equity table
and jump to 2-away/3-away directly, without training separate
nets for the smaller scores.) And then obviously, you can
work your way up the chain of match scores, training separate
nets for each score.
if you take into account what a amount of resources deepmind
throws in, they would surely come out with something better,
but not washing the floor with any of XG, BGBlitz or GnuBG.
The alphago team was 18 people.... Compare this to leisure
time programmers (BGBlitz, GnuBG) or a single full time
programmer (Xavier?).
deep mind always tries to do thing that haven't been done yet.
Unless one of their Cxx get's mad on BG they wont do BG.
Again, cubeful or cubeless, I'm not so sure that N-away/N-away
is the same in any match length. I'll need to digest this and may
need to be convinced whether by others or by myself.
On Friday, December 24, 2021 at 6:11:16 AM UTC-5, MK wrote:
Again, cubeful or cubeless, I'm not so sure that N-away/N-away
is the same in any match length. I'll need to digest this and may
need to be convinced whether by others or by myself.
I'm not sure I understand exactly what you're saying here.....
On the other hand, if you're questioning whether a score of 2-2 in
a match to 5 points is the same as a score of 4-4 in a match to
7 points---they are the same. The history of what has happened
before in the match cannot influence your strategy; you are now
playing a match to 3 points and have to win it.
You said that you would be satisfied with training something
to play a single cubeful money game, but what I'm suggesting
is that the way to approach it is via match play.
I mentioned before some of my colleague's preliminary
experiments.
If you try to do this kind of training directly, the unlimited cube
values will make it very difficult for the neural net to learn from experience.
I tried to explain this to you before, but if you don't believe me,
the only way to convince yourself is to try it and see for yourself.
Am I wrong to think that cubeful money game is what would
bring out the best/most out of the players' "cube skills", if there
is indeed such a skill??
May be difficult but not impossible, yes? You can tacke this in
small, gradual steps just as you suggested for various N-aways
in match play. Start with limiting the cube at low values at first.
Then gradually increase them and try get as close as you can to
unlimited cube values.
I already did convince myself. But there is no way for me to
convince you people. That's why I'm trying to make you all
convince yourselves.
May be difficult but not impossible, yes? You can tacke this inThat is almost exactly what I was proposing---start with lower
small, gradual steps just as you suggested for various N-aways
in match play. Start with limiting the cube at low values at first.
Then gradually increase them and try get as close as you can to
unlimited cube values.
match scores where there is a natural cap on the size of the cube,
and work your way up.
- There's no reason to bring DeepMind into this discussion. The basicBecause many people in similar discussion expressed the belief that if Deepmind will do BG they will wash up the floor with the current bots, just because they learned chess in a couple of hours (on a fast PC it was about a year of training, if you don't
ideas are out there, and other people can more or less replicate the results, with much less effort than was needed to come up with the ground-breaking ideas in the first place.
winning rate from the starting position. Surely it's obvious that theI don't agree completely on that. You see that here GnuBG is completely off whereas BGBlitz plays reasonable ( http://bgblitz.com/olympiad_2016.html ) and on Dailygammon (dailygammon.com Discussion search for: threads like Position #25 that are posted
point is to understand superbackgames and wild containment games and
other positions which the current bots have less than zero clue about.
On December 21, 2021 at 4:12:07 PM UTC-7, Frank Berger wrote:Too bad we won't see it. I'm quite sure that Deepmind would only invest resources, if they see a possibility for a breakthrough.
if you take into account what a amount of resources deepmindOn the opposite end, I bet that an alpha-zero-bg-bot would run
throws in, they would surely come out with something better,
but not washing the floor with any of XG, BGBlitz or GnuBG.
XG and GnuBG through the garbage disposer. (BTW: BGBlitz is
not in the same class as the other two.)
I'm not sure. there are a lot of small tweaks to improve and not every idea does work, so to competet it might take some work.The alphago team was 18 people.... Compare this to leisureAll notable BG bots since TD-gammon v2 are offshoots of it.
time programmers (BGBlitz, GnuBG) or a single full time
programmer (Xavier?).
So, it's easy for even a single amateur/leisure programmer or
a petty-crook/professional programmer to copy and release
a slightly improved version of a previous BG bot.
Personally, I think "they" are avoiding BG because of how it'sAs I mentioned, they probably don't see a real breakthrough. Let's say that they do a bot that has a PR of -0.5. If they do a real match and they don't have a playing style to do backgames all the time and confude the opponents they may easily loose (
*infested* by gamblers.
posted by Zorba) a very deep backgame was discussed intensively recently. In some positions XG doesn't find a reasonable move whereas BGBlitz founds it with 1-5 ply. Not always that extreme but for the small sample I have (please send me extremewinning rate from the starting position. Surely it's obvious that theI don't agree completely on that. You see that here GnuBG is completely off whereas BGBlitz plays reasonable ( http://bgblitz.com/olympiad_2016.html ) and on Dailygammon (dailygammon.com Discussion search for: threads like Position #25 that are
point is to understand superbackgames and wild containment games and
other positions which the current bots have less than zero clue about.
XGID=-eEeEeE-------------------:1:1:1:00:0:0:0:0:10not yet. It does understand to roll home the prime, but it does not yet understand to slot at the front of the prime. You may try it yourself.
Tim Chow schrieb am Donnerstag, 30. Dezember 2021 um 14:33:11 UTC+1:
XGID=-eEeEeE-------------------:1:1:1:00:0:0:0:0:10not yet. It does understand to roll home the prime, but it does not yet understand to slot at the front of the prime. You may try it yourself.
And between getting the most extreme position perfect and totally lost is a continuum. At least my definition of totally lost is not met, your mileage may vary.
And at the Kauder paradox we as Humans have an edge but what about XGID=---a--A--BBB--ABa---BbAaAA:0:0:1:41:5:6:1:7:10 ? XG has, after 15 hours of rollouts b-24,10-6 as first and the obvious(?) b-24,20-16* only on pos 6, -0,235 worse. BGBlitz gotthis move right on 1-5-ply. (see the discussion and more data here: http://www.dailygammon.com/bg/forum2/main/read/56208 ) . But I think this not so difficult for humans.
But what about that XGID=---A-Aa--BBCA-----A--d-BB-:0:0:1:21:5:6:1:7:10 ? I wouldn't trust anyone who "knows" what is right here....
But what about that XGID=---A-Aa--BBCA-----A--d-BB-:0:0:1:21:5:6:1:7:10 ? I wouldn't trust anyone who "knows" what is right here....This is my point. There's still plenty of room for improvement
in backgammon if you allow for "prop" play. Choose a position
and play 100 games as White and 100 games as Black, and see who
comes out ahead. It's totally possible that a new bot could
crush current bots under those circumstances.
If you start with a money game, but then impose a
limit on the cube, that is almost exactly the same
as playing a short N-away/N-away match... So once
you start talking about capping the cube, your
proposal is not materially different from mine.
On Sunday, December 26, 2021 at 12:28:24 AM UTC-5, MK wrote:
Am I wrong to think that cubeful money game is what
would bring out the best/most out of the players'
"cube skills", if there is indeed such a skill??
The conventional view is that match play is a better
test of cube skill because the cube handling varies
with the match score.
Start with limiting the cube at low values at first.
Then gradually increase them and try get as close
as you can to unlimited cube values.
That is almost exactly what I was proposing---start
with lower match scores where there is a natural
cap on the size of the cube, and work your way up.
I already did convince myself. But there is no way
for me to convince you people. That's why I'm trying
to make you all convince yourselves.
The reason you can't convince us is because you
operate outside the realm of facts and logic. You
convince yourself about how AlphaZero would
perform without understanding the first thing
about deep learning. Of *course* you can't convince
others who actually know something about the subject.
Personally, I think "they" are avoiding BG because of how
it's *infested* by gamblers.
As I mentioned, they probably don't see a real breakthrough.
Let's say that they do a bot that has a PR of -0.5. If they
do a real match and they don't have a playing style to do
backgames all the time and confude the opponents they
may easily loose (dice, you know). That doesn't seem a
good invetsments for a 6 to 7 digit investement.
I don't understand the "gambler" thing. BGBlitz learns
through selfplay, I don't know where there is influence.
I'll do that in 2022 with a freshly breeded AI (March? April?
.....
The good thing is: we know the equity numbers exactly
Too bad that so few people are interested in BG-AI compared
to e.g. chess :( but probably there are some reasons for it.
Debating with you
gets as tiring and useless as arguing with creationists
who insist that dinasours could fit in Noah's Ark; well
at least baby dinasours or maybe even dinasour eggs...
On 1/5/2022 11:47 AM, MK wrote:
Debating with you gets as tiring and useless as
arguing with creationists who insist that dinasours
could fit in Noah's Ark; well at least baby dinasours
or maybe even dinasour eggs...
So why do you do it?
I debate with because I find you hilarious. But you
don't find me hilarious.
On January 5, 2022 at 7:37:02 PM UTC-7, Tim Chow wrote:
On 1/5/2022 11:47 AM, MK wrote:
Debating with you gets as tiring and useless as
arguing with creationists who insist that dinasours
could fit in Noah's Ark; well at least baby dinasours
or maybe even dinasour eggs...
So why do you do it?
Partially because others read what I write to you here.
Why do you think you matter that much to anybody??
BTW: was Axel your secret colleague who had taken a
shot at developing a Alpha-Zero Hypestgammon bot?
Sysop: | Keyop |
---|---|
Location: | Huddersfield, West Yorkshire, UK |
Users: | 296 |
Nodes: | 16 (2 / 14) |
Uptime: | 78:31:50 |
Calls: | 6,658 |
Calls today: | 4 |
Files: | 12,203 |
Messages: | 5,332,975 |
Posted today: | 1 |