From trains to planes...
Murat Kalinyaprak
After take off, traveler1 pulls out his laptop
and starts playing a computer backgammon game.
Then the following exchange takes place:
Player2: Can I play a game against it too?
Player1: How much money do you want to play for?
Player2: I'll let you pick the stakes, if you...
Player1: (Hastily interrupts) Do you have any
idea what this program is called?
Player2: No, and that's why I let you pick the
stakes, if you let me pick the seed and
counter...
(No offence intended against anyone... :)
MK
my_...@hotmail.com
Dear golf enthusiasts,
I have grave suspicions that world-renowned Tiger Woods
cheats. And on top of that, I can also prove he is not
a world class golfer. I feel that people defending TW
on this subject must be either taking a cut from his
revenues, or must be a little naive.
Firstly, I invited Tiger over to my house for a putting
contest. After 100 putts, my score was 2% better than
his. And he calls himself a golfer? Obviously I do not
claim to be world-class calibre, but if I can continue
to beat this mighty "Tiger Woods" by a 2% margin during
putting, surely I can thrash everyone at professional
golf tournaments!
Although I did not pay Mr. Woods for his time, I feel
it necessary to complain publicly about him. Further-
more, I sensed through telepathy that there's something
very wrong with the "random ball dispenser" on his golf
bag. I believe I can predict exactly what brand of ball
will pop out. That, surely, gives Mr. Woods an enormous
advantage. As well, one time I went up to his dispenser
and uttered the magical phrase "777777777", and instead
of a golf ball, out came a pine cone! I tried to hit it
and it travelled a negative distance!!! I gave the pine
cone to Tiger and the same thing happened!!!! I was so
disgusted by this that I wrote a letter to his father
Fredrik demanding an explanation. Fredrik told me that
raising a child to play the sport of golf very well was
extremely difficult, and that occasionally a few pine
cones get left in golf bags. However, I did not under-
stand this reasoning so I made Fredrik repeat himself.
One person told me there are lots of reasons to suggest
that Tiger Woods is a good golfer; his golf rating from
playing 6000+ holes, his peers saying that he is indeed
a worthy golfer, young children who imitate his style
and want to grow up to be just like him..but never mind
that. I think Tiger's biggest weakness is his inability
to "play the man" and alter his swing to compensate for
his opponents'. Tiger merely tries to play each shot
optimally, while as we all know, one should change ones
game as much as possible from day to day.
Keep in mind that I have nothing against this deceitful
and fraudulent Tiger Woods. I merely wish to learn the
truth, and don't we all? Thanks, and have a good day.
-----== Posted via Deja News, The Leader in Internet Discussion ==-----
http://www.dejanews.com/rg_mkgrp.xp Create Your Own Free Member Forum
Murat Kalinyaprak
> Dear golf enthusiasts,
> I have grave suspicions that world-renowned Tiger Woods
> cheats. And on top of that, I can also prove he is not
> a world class golfer. I feel that people defending TW
> on this subject must be either taking a cut from his
> revenues, or must be a little naive.
Although you use spaces to align text (which may be
considered "cheating" by some:), I appreciate your
efforts, along with your point-by-point analogies
which I like, and think they deserve a response.
I don't mind sarcasm either but one must be careful
that it doesn't end up cutting both ways...
> Firstly, I invited Tiger over to my house for a putting
> contest. After 100 putts, my score was 2% better than
> his. And he calls himself a golfer? Obviously I do not
> claim to be world-class calibre, but if I can continue
> to beat this mighty "Tiger Woods" by a 2% margin during
> putting, surely I can thrash everyone at professional
> golf tournaments!
Just to be picky, I think your 2% is wrong and your
analogy would have worked better if you used a
computer golf game instead. The important point is,
I never claimed that I could trash everyone in
professional backgammon tournaments by measuring
myself against JF. In fact, it's the other way
around. That is, I constantly expressed that JF may
not be a good measuring stick and that I believed
I couldn't do nearly as good against other humans
(worl-class players) which JF supposedly equals or
beats.
> Although I did not pay Mr. Woods for his time, I feel
> it necessary to complain publicly about him.
As I was about the erase it from my hard drive,
after a few days of trying it, I happened to
express my opinion about it (which as coming from
a "nobody" who had never been heard of before, it
shouldn't mean much anything to anybody anyway.
Claes Thornberg
> Dear golf enthusiasts,
> I have grave suspicions that world-renowned Tiger Woods
> cheats. And on top of that, I can also prove he is not
> a world class golfer. I feel that people defending TW
> on this subject must be either taking a cut from his
> revenues, or must be a little naive.
Although the writer of this inane post has grasped the only
interesting thing with Murat's posts, I must say his effort
to write perfectly blocked text is crippled. By using space
characters to pad the lines to the right length, he's shown
his lack of linguistic talent, which is required to succeed
in this area. Before posting again, I suggest spending some
time reading earlier posts with perfectly blocked text.
Regards,
Claes Thornberg
PS. Yes I know the post I am commenting was supposed to be,
even though I didn't find it so, funny.
--
______________________________________________________________________
Claes Thornberg Internet: cla...@it.kth.se
Dept. of Teleinformatics URL: NO WAY!
KTH/Electrum 204 Voice: +46 8 752 1377
164 40 Kista Fax: +46 8 751 1793
Sweden
Ian Shaw
not knowing anything about statistics, RNGs, Neural Network Development, or
even doubling and tournament play.
From his point of view, he downloaded a BG program which is rated as
world-class, and found he could beat it. He then started asking questions.
Some of them seem daft to those more accomplished in the relevant fields,
and some have shown a fair degree of lateral thinking.
He has responded positively to all criticism. He has correct his erroneous
assumptions about the measurement parameters once details of the various
scoring systems were explained, and still achieved decent scores in single
games (which is the only type of BG he is familiar with).
He has also found a bug. As a programmer, I know it is not serious;
nevertheless, a bug is a bug is a bug. As a non-programmer, Murat is quite
entitled to speculate wildly on the possible consequences of this bug.
The upshot of it all is this:
I'd be very interested to know what people think is a good way of evaluating
ones skill level against JF, in a reasonable number of games. Surely there
must be some test that does not involve thousands of games to identify a
slight edge, or what is the point of all the tournaments?
Regards,
Ian
Of course, at Microsoft, a bug is not a bug, its an issue. I suppose this is
an example of the Seattle sense of humour. If one catches a bug, one goes
a-issue, a-issue...
Claes Thornberg
>
> I think Murat deserves to be cut a bit of slack. It is OK to play BG whilst
> not knowing anything about statistics, RNGs, Neural Network Development, or
> even doubling and tournament play.
Yes fortunately it is. But when discussing whether or not Jellyfish
cheats, you have to discuss it in terms of statistics, RNGs, and NN
theory. People have tried to explain these concepts but the problem
is that he doggedly sticks to his opinions without bothering trying
to understand what other people are writing.
[snip]
> He has also found a bug. As a programmer, I know it is not serious;
> nevertheless, a bug is a bug is a bug. As a non-programmer, Murat is quite
> entitled to speculate wildly on the possible consequences of this bug.
And speculating wildly he has surely done!
> The upshot of it all is this:
> I'd be very interested to know what people think is a good way of evaluating
> ones skill level against JF, in a reasonable number of games. Surely there
> must be some test that does not involve thousands of games to identify a
> slight edge, or what is the point of all the tournaments?
Unfortunately I do not think there is any other way. The closer you
are to Jellyfish in playing strength, the more games you'll need to
play. This is a consequence of statistics. Then what's the point of
all the tournaments? The easy answer is that people like to gamble.
A better answer is that the good players might not succeed in every
tournament, but in the end they come out a head of almost everyone.
> Regards,
> Ian
Regards,
Claes Thornberg
Ian Shaw
Claes Thornberg wrote in message ...
>"Ian Shaw" <ian....@riverauto.co.uk> writes:
>
>>
>> I think Murat deserves to be cut a bit of slack. It is OK to play BG
whilst
>> not knowing anything about statistics, RNGs, Neural Network Development,
or
>> even doubling and tournament play.
>
>Yes fortunately it is. But when discussing whether or not Jellyfish
>cheats, you have to discuss it in terms of statistics, RNGs, and NN
>theory. People have tried to explain these concepts but the problem
>is that he doggedly sticks to his opinions without bothering trying
>to understand what other people are writing.
>
I think Murat is now pretty much convinced that JF is not deliberately
rigged. He is now contemplating whether JF during its training process could
have learnt its strategy based on some pattern in the rolls, which no one
has spotted yet. You've got to admit he thinks creatively. From reading the
posts, I've been left with the impression that Murat is quite willing to
change his views if he is prented with a good case.
What really ticked me off was the anonymous flame.
I am of the opinion that the RNG bug is evidence that JF is *not* programmed
to cheat. If programming a foolproof RNG is beyond Fredrik (no offence
intended), then what chance has he got of implementing a highly complex,
undetectable cheating system. It's just too difficult.
Have any Scrabble enthuiasts run a Murat's name through an anagram
generator? Just stirring the pot :) ...
Ian
Rodrigo Andrade
text?
RODRIGO
Murat Kalinyaprak
half of the article.
my_...@hotmail.com wrote:
> Dear golf enthusiasts,
> I have grave suspicions that world-renowned Tiger Woods
> cheats. And on top of that, I can also prove he is not
> a world class golfer. I feel that people defending TW
> on this subject must be either taking a cut from his
> revenues, or must be a little naive.
Although you use spaces to align text (which may be
From older articles, it seemd like this had become
a routine which was countered by stock-responses
based on seed/counter, etc.) Since JF was available
for free evaluation (and practically free use if
one was willing to put up with the nag screens)
nobody had to consider my opinion and could try it
out for themselves. So, it wasn't a question of
"complaining" and I got further into these
discussions just because I felt it wasn't right
that people were scolded/put down for having
expressed their impressions/opinions about JF.
> Further-
> more, I sensed through telepathy that there's something
> very wrong with the "random ball dispenser" on his golf
> bag.
There was indeed a problem and more may or may not
be discovered in time. What you call "telepathy"
was what I thought to have "observed" (i.e. games
against JF not feeling real-life like, etc.). Since
I myself often volunteered that it may have been
just my "imagination", you are free to discard my
observations by calling them whatever you like.
> I believe I can predict exactly what brand of ball
> will pop out.
Not quite "exactly" but felt I could do often enough
to believe that I could use it to my advantage in
battling against JF.
> That, surely, gives Mr. Woods an enormous advantage.
Not "surely", does but "could"... And the opposite
is possible also. I expressed many times that if
indeed there were some detectable patterns, they
could very well be exploited against JF...
> As well, one time I went up to his dispenser
> and uttered the magical phrase "777777777", and instead
> of a golf ball, out came a pine cone!
You have good imagination and I find this funny. How
would you have felt if this really happened? :) It
appears that some of the articles I posted never made
it to other newsservers, but I had given more examples
of this bug in another article, with other seed/counter
settings. Plus, I have observed other different
manifestations of this bug, which I haven't yet
mentioned but I may later if I deem that they may be
further useful.
> I tried to hit it
> and it travelled a negative distance!!!
Yes, indeed...
> I gave the pine
> cone to Tiger and the same thing happened!!!! I was so
> disgusted by this that I wrote a letter to his father
> Fredrik demanding an explanation.
Who said I was "digusted"? I was rather amused by it.
Don't you think that it was better for others and
expecially Fredrik to know about it, so that he can
correct it in future revisions of his program? I
haven't "demanded" anything either. Initially I had
expressed that he sould be given a chance and "enough"
time to validate there was a bug and explain (if he so
wished) how/why this was happening. And yes, later I
found his response inadequate. Fredrik may be too busy
or whatever to bother with lengthy clarifications but
it seems like you have the spare time, and maybe you
can answer the questions I posed/(implied). Can you?
> Fredrik told me that
> raising a child to play the sport of golf very well was
> extremely difficult, and that occasionally a few pine
> cones get left in golf bags. However, I did not under-
> stand this reasoning so I made Fredrik repeat himself.
I'm of the opinion that checking for invalid dice
numbers would be a trivial/elementary part of any
dice generating algorithm. I had added to that, the
fact that I had played 100 games each with the seed
set at 999999999 and 888888888, having rolled 5000+
dice rolls during each set of 100, but not observing
anything unusual. But with seed settings like
777777777, 888888887, etc. several occurences of this
bug happen righ after a few rolls. Now, if you could
sit down and write a dice generating algorithm which
emulates this bug, you could really ridicule me
beyond these funny riddles. Do you think you can
do it...?
> One person told me there are lots of reasons to suggest
> that Tiger Woods is a good golfer; his golf rating from
> playing 6000+ holes, his peers saying that he is indeed
> a worthy golfer, young children who imitate his style
> and want to grow up to be just like him..but never mind
> that.
I haven't had a chance to explore whether JF plays
better with its own dice vs. manual dice, but I had
expressed from early on that I think it plays quite
strongly even with manual dice. I have no problem
with anybody's rating out there, including JF's. I
just have a problem with "my own unexpected rating"
against JF...
> I think Tiger's biggest weakness is his inability
> to "play the man" and alter his swing to compensate for
> his opponents'.
Yes, I happen to believe this is the stregth of
humans and the weakness of machines. A human may
achieve positive results against JF which will be
visible rather quickly but a similar accomplishment
on JF's part may take thousands of games before
becoming visible by meeting expectation based on
large numbers of statistics. I'm suprised that so
many people can't grasp this concept/argument.
> Tiger merely tries to play each shot
> optimally, while as we all know, one should change ones
> game as much as possible from day to day.
If a human opponent "figures you out" after a few
games and you can't do the same, you may be in
bigger trouble than you would ever be against JF.
JF will never be able to "throw you out" because
you already know (more or less) how it will play
forever (an it will take it close to forever to
"even out the odds":).
> Keep in mind that I have nothing against this deceitful
> and fraudulent Tiger Woods. I merely wish to learn the
> truth, and don't we all? Thanks, and have a good day.
In my previous articles, I elaborated on how the
term "cheating" may be relative/ambiguous. In
short, even the effects of unintentional flaws
may be considered "cheating" for all practical
purposes from the recipient's (user's) point of
view. If you still have a problem with this
approach, I can't help it. I guess it's everyone's
own business how much they want to know about what.
MK
Murat Kalinyaprak
> I think Murat deserves to be cut a bit of slack......
Ian, thanks for being willing to extend this. I found your
remarks (which I haven't duplicated here) reasonably
fair and accurate also.
> He has also found a bug. As a programmer, I know it is not serious;
> nevertheless, a bug is a bug is a bug. As a non-programmer, Murat is quite
> entitled to speculate wildly on the possible consequences of this bug.
I guess time has come for me to mention that I have
18+ years experience as a programmer/analyst. For the
many years that I have been on the Usenet, I tried to
participate in discussions without bringing in such
parameters as age, degrees, marital status, etc. into
the subject unnecessarily. But in this case, if I had
witheld this any further, I thought I may end up being
accused of sand-bagging...
Game programs, random number generators, etc. have not
been my areas of specialty, but I hope that I have by
now some general ideas about how numbers are stored
internally, etc. If this bug is as trivial as it seems
to you, so much the better especially for Fredrick who
I suppose would eventually want to fix it. However, as
a result of experience or habit, I don't take bugs for
granted.
I don't know if my speculations are wild or not, but
I had questioned the size of JF's seed/counter values
and it seems like it has now been validated that I was
not completely unjustified to question it. There are
of course more questions remaining, like: if JF itself
was checking/making sure it didn't pick seed values
(each time you start it, based on system clock) greater
than 65535, why was the user allowed to enter a 9 digit
value? Or about the counter, for example, why is the
user is limited to entring a 5 digit value while JF
itself increments it well past that (I suspect it would
go all the way up to 2 or 4 billion)? Whether they
expressed them openly or not, I would like to think
that most people who followed or participated in these
threads must have already asked these or other similar
questions themselves.
> The upshot of it all is this:
> I'd be very interested to know what people think is a good way of evaluating
> ones skill level against JF, in a reasonable number of games. Surely there
> must be some test that does not involve thousands of games to identify a
> slight edge, or what is the point of all the tournaments?
Statistics span large numbers of games/dice rolls and
given time I'm sure that strategies based on them will
eventually produce positive results. But I believe
that strategies based on pure statistics can be defeated
in the short term, not just by pure luck but by skill.
I'm not sure if 50 or 100 games are enough to be
considered as "telling", but I sure don't think that
2000 games or 2 billion dice rolls are necessary either.
If there is any claim that there is room for something
called "strategy" in backgammon, then it must be accepted
that humans may be capable of devising/adopting some
"non-statistics based strategies" to achieve rather quick
results.
MK
Murat Kalinyaprak
> Claes Thornberg wrote:
> I think Murat is now pretty much convinced that JF is not deliberately
> rigged.
Not "deliberately" in the meaning that with "bad intentions".
To repeat myself, I have expressed from early on that JF still
plays well with manual dice. The only purpose remaining in
these discussions is whether people who express their opinions
about JF's "apparent luck" deserve being scolded/put down or
whether they may be actually observing something. Notice that
from the very beginning, I never used the term "lucky" for JF.
I had realised very quicly that advancing the counter by one
number manually, the same rolls would go the opposite player.
So, the questions related to it become things like, if there
is an anomaly in the dice rolls (i.e. unusually frequent
doubles for example), can it be possible for a machine to
take better advantage of them than a human can, etc.
You may wonder what I would consider "deliberately rigged
with good intentions"? For example, I had previously raised
issues like whether the same roll could happen 5, 6, 7 times
one after another. If it happened, would JF allow it or would
it try curb it by allowing only 4 of them and supressing the
rest by a filtering algorithm? Personally, I wouldn't call
any such attempts (to make dice rolls more even and less cruel
than pure random) necessarily cheating. And such attempts may
indeed make the entire task of dice rolling complicated to
the degree of causing inadvertent anomalies/bugs. Nevertheless,
the fact that it may be inadvertent doesn't take away the
user's right to perceive this as causing them disadvantage.
> He is now contemplating whether JF during its training process could
> have learnt its strategy based on some pattern in the rolls, which no one
> has spotted yet.
I think this had already been ruled out by the argument that
JF was trained using a different dice rolling algorithm...(?)
> You've got to admit he thinks creatively.
Among all the heat I receive, such comments are as welcome
as a cool breeze... Thanks and I think you (and couple of
other writers) are leading me write in a more positive tone
than I use answering some people's smart remarks.
> From reading the posts, I've been left with the impression that
> Murat is quite willing to change his views if he is prented with
> a good case.
Definitely. Actually, I continue these discussions for the
sake of "logical thaught process" more than anything else.
I couldn't care less if they were about Jellyfish, Snowie
or some computer bingo game.
I don't know if readers realize it but what opens the door
for suspicions/speculation/discussions on this suject and
creates/perpetuates a "challenge" is the fact that the
author of this program has gone as far as creating/offering
a seed/counter mechanism in order to prove that JF doesn't
cheat. The problem is that the seed/counter argument alone
doesn't cut it. I have no idea why he hasn't made the dice
rolling algorithm public instead of or in addition to that
(which I believe would have put an end to such discussions
once for all, a long time ago).
It looks like a lot of arguments have been and are still
being made to prove that JF's playing ability has nothing
to do with its dice rolling and I think they are valid.
In addition to that, I have never seen any arguments made
about JF's being a "world-class dice roller". Obviously,
its dice rolling is inconsequential to its real value. It
also seems like very good dice rolling algorithms are in
the public domain already. In light of these, personally
I can't come up any reasons why I wouldn't have made it
public, if I were the author of a similar program. In a
previous article, I mildly suggested that Fredrik may
want to consider this. Whether he does or not is up to
him. But as long as there is something kept from them
(for seemingly no good reason) people will wonder, suspect,
speculate, etc. It's just human nature...
> Have any Scrabble enthuiasts run a Murat's name through an anagram
> generator? Just stirring the pot :) ...
Oh, oh...
MK
Murat Kalinyaprak
> "Ian Shaw" <ian....@riverauto.co.uk> writes:
>> I think Murat deserves to be cut a bit of slack. It is OK to play
>> BG whilst not knowing anything about statistics, RNGs, Neural
>> Network Development, or even doubling and tournament play.
> Yes fortunately it is. But when discussing whether or not Jellyfish
> cheats, you have to discuss it in terms of statistics, RNGs, and NN
> theory.
But you need to do that specificly relating to JF, not
just in general. You guys put me through RNG-101 here,
in general, yet I'm not sure if any of it related to JF
beyond the argument: "if 32 bit "space" (Gary's term?)
is already there, why not use it?". As it happens now,
it seems like if JF's seed values go past 16 bit (i.e.
65535), it causes problems. What now? I acknowledge that
I'm not an expert on RNG's, NN's, etc. but can you even
address the issues I'm raising in my limited capacity,
before you go on constantly harping on my lack of deeper
knowledge in these subjects?
> People have tried to explain these concepts but the problem
> is that he doggedly sticks to his opinions without bothering trying
> to understand what other people are writing.
I appreciate your responding to my articles in lebgth,
but I don't understand why you keep repeting this? Yes,
I haven't yet been able to prove/disprove whether there
is any basis for what I believe to observe. But how can
I do that, without for example clicking on the mouse for
six months to accumulate enough data to analyze and come
to a conclusion one way or another? I already offered
that it wouldn't mind making some effort (even with the
aim of disproving my own impressions) if it was practically
feasible/convenient. So, I'm not really sticking to any
specific "opinions" but only insisting that some questions
remain unanswered. I just posted some numbers about the
average game lengths and average dice values rolled in a
certain amount of games I had played against JF. What did
you think of them? Do you yourself have any better stats
showing that there are no unintentional flaws in JF's
dice generator...?
>> He has also found a bug. As a programmer, I know it is not serious;
>> nevertheless, a bug is a bug is a bug. As a non-programmer, Murat is quite
>> entitled to speculate wildly on the possible consequences of this bug.
> And speculating wildly he has surely done!
I don't think I speculated anything on this particular
issue yet. I simply indicated the possibility that it
may not be as trivial as it seems to some readers and
left it at that.
>> The upshot of it all is this:
>> I'd be very interested to know what people think is a good way of
>> evaluating ones skill level against JF, in a reasonable number of
>> games. Surely there must be some test that does not involve thousands
>> of games to identify a slight edge, or what is the point of all the
>> tournaments?
> Unfortunately I do not think there is any other way. The closer you
> are to Jellyfish in playing strength, the more games you'll need to
> play. This is a consequence of statistics.
And I find it *unfortunate* that backgammon playing is
reduced to mere statistics. After that, what's left for
the human brain? It just becomes a matter of your finger
clicking on the "show roll-out", "calculate equity",
"move piece" icons and nothing more. It completely does
away with the concept of "strategy" and frankly I think
it kills the "spirit" of the game...
MK
Claes Thornberg
Yes, but if you think a little, you can manage yourself without using
pad characters. By rephrasing what you are going to say, or by adding
commas which don't alter the meaning of what you are writing, you can
write blocked text quite easily. My own method, and I haven't been at
this for so long, is just to start writing. My word processor usually
breaks the first line making it between 60 and 70 characters long. By
setting the right margin appropriately, I won't get lines longer than
that, and when seeing that a line is getting too short, or long, it's
often easy to change some words, make or resolve contractions. Simply
write "I am" instead of "I'm", "it is" instead of "it's", etc. Once I
got the hang of this, I found it hard to stop. Just keep in mind, the
longer lines you make, the easier it is to write blocked text without
pad characters.
Laury Chizlett
>I think Murat deserves to be cut a bit of slack. It is OK to play BG whilst
>not knowing anything about statistics, RNGs, Neural Network Development, or
>even doubling and tournament play.
>
>world-class, and found he could beat it. He then started asking questions.
>Some of them seem daft to those more accomplished in the relevant fields,
>and some have shown a fair degree of lateral thinking.
>
I agree. I dont know about Murat's experience, but I believe that BG
players tend to play in small groups of other players, and get very
little input from any other groups. This is a result of the
comparatively low number of BG players compared to Chess, say. It is
very rare to see a BG column in a newspaper, for example. There is no
world-wide organisation (as far as I know).
My particular group consists of my father (until he died), and two
school friends, and now my wife. I had to teach all of these. I had
never found a book, and had no idea that anyone would ever write one.
Our knowledge evolved like Darwin's animals on the Galapagos Islands:
totally isolated from the rest of the world.
Examples of this parallel evolution were
- fairly early on we found the advantage of the advanced anchor,
only we called it a "safety"
- blitzes were called "tsunami" (tidal or harbour waves)
- primes, we called "blocks"
When I got on the Net, one of the first things I did a search on was BG
(I will never know why) and was suddenly plunged into another world of
take points, jellyfish, mandatory doubles, equity and (my favourite)
recube vigorish. It was a hell of a shock but, ultimately, a good shock.
I too play Jellfish, level 7, in 1 point matches and 5 point matches. I
win 70% of the former (If I think straight) and about 50% of the latter.
I dont think it cheats: JF once got 14 men off against me in a double-
match point game, when I hit with a full home board. JF then proceded
throw doubles - only trouble was they were ones and twos. When I had one
man left on the ace point, he woke up and hit me. He then dealt me a 66
to win.
-- ^ To Liverpool St
^ Station & City
Laury | ^ | 1.5 miles
________________________| |
TRP Ridley Rd Street Market |__
35 Colvestone Crescent __________________ _ | | Dalston
London / / | |__| Kingsland
E8 2LG / / |A10 | Station
________________/ / | |
Tel: 0171 923 0244 Colvestone Cres / | |
Fax: 0171 923 1471 ____________________/ | |
35
Rew Francis
Eibisch) wrote:
>Yes, don't I
>attempt to did
>block text think
>this thin, of this
>unless you shape for
>enjoy word an article,
>play. Some but then this
>days ago I group is really
>did a post about backgammon,
>limited to not text tricks
>a scant 12 like this. As
>characters a rule keep
>and it was r.g.b. to
>reasonably ontopic
>difficult. stuff
>This one's say
>10 letters I
>wide and I
>am writing
>it quickly
>as you can
>tell!
>
>--
>James _ To mail me, spell "nomed" in my address backwards
>Eibisch, ('v')
>'Ivan' (,_,) N : E : T : A : D : E : L : I : C : A
>on FIBS. ======= http://www.revolver.demon.co.uk
I did
think
James
might
write
posts
which
tried
out a
block
of at
least
10 or
so- I
can't
see 5
being
worth
one's
while
if it
makes
for a
style
which
seems
a bit
fake.
Rew
Murat Kalinyaprak
>James Eibisch wrote:
>>Yes, don't I
>>attempt to did
>>block text think
>>this thin, of this
>>unless you shape for
>>enjoy word an article,
>>play. Some but then this
>>days ago I group is really
>>did a post about backgammon,
>>limited to not text tricks
>>a scant 12 like this. As
>>characters a rule keep
>>and it was r.g.b. to
>>reasonably ontopic
>>difficult. stuff
>>This one's say
>>10 letters I
>>wide and I
>>am writing
>>it quickly
>>as you can
>>tell!
>I did
>Rew
For being someone who can block 5 character
long lines, wouldn't you wish your name was
something like James, Murat or Steve...? :)
I feel good that I could start a fun fad in
this group, but I had no idea that it would
get this far. You guys are just amazing...!
MK
Dan Frank
Murat Kalinyaprak <mu...@cyberport.net> wrote
Subject: Re: What is a fair test of ability?
Statistics span large numbers of games/dice rolls and
given time I'm sure that strategies based on them will
eventually produce positive results. But I believe
that strategies based on pure statistics can be defeated
in the short term, not just by pure luck but by skill.
I'm not sure if 50 or 100 games are enough to be
considered as "telling", but I sure don't think that
2000 games or 2 billion dice rolls are necessary either.
If there is any claim that there is room for something
called "strategy" in backgammon, then it must be accepted
that humans may be capable of devising/adopting some
"non-statistics based strategies" to achieve rather quick
results.
*********************************************************
Until now I didn't read MK's postings: I'm not interested in that matter
and they are toooooo long. Yesterday as I was marking text from a down
loaded Usenet file to delete it - by chance I've noticed the upper text.
I can subscribe to it (up to some details, where I haven't made my
mind up yet), in fact, that's another way of expressing what I think,
see f.i.
Moves are not good, because figures are good, but figures are good
because moves are good!
That's the prevalence of strategy over statistics, of quality over quantity ASF.
--
Dan Frank
editor & publisher of ESSENTIAL BACKGAMMON
Marina Smith
Kalinyaprak) wrote:
>For being someone who can block 5 character
>long lines, wouldn't you wish your name was
>something like James, Murat or Steve...? :)
Marina For another Then, if you had
is not column, you your Thesaurus
so bad can provide to hand, you
either a different could do a
if you width for a tasteful
can be longer name shaped
blithe - fibs name bit.
with 6 or possibly
letter a different
words. server sig.
Marina mas on fibs
--
Marina Smith - Reading, UK. To email me, remove XX from my address.
Chuck Bower
Ian Shaw <ian....@riverauto.co.uk> wrote:
>I think Murat deserves to be cut a bit of slack. It is OK to play BG whilst
>not knowing anything about statistics, RNGs, Neural Network Development, or
>even doubling and tournament play.
True, but I got on his case NOT for these reasons, but for breaking
my (self-imposed and admittedly subjective) "rules of r.g.bg civility":
1) Profanity is INEXCUSIBLE. (Murat has NOT broken this rule, but...)
2) Strong accusations (such as of cheating or 'dropping') require strong
evidence. (See Murat's post of July 8 for his intial comments.) See his
many other posts for his failure to provide the evidnece. (No, finding
a bug in the RNG input routine is not sufficient evidence of cheating!!!!!!)
3) Insincerity and hypocrisy are out of place here (and everywhere else on
the face of the earth, for that matter). This one can be tough to judge
because often people make 'tongue in cheak' comments (and scarcasm) which
are subject to misinterpretation, for example. But feigning modesty
with an ego that doesn't fit through a canyon is insincerity, in my opinion.
Put 2) and 3) together and you reach threshold in my book for getting
a harsh reply post from me. (And I'll bet THAT scares you!)
I've been asked why I bother with these kind of posts and replies,
since it takes time away from seemingly more valuable work (like analyzing
positions). Similarly, I've asked some others why they DON'T reply to
unsubstantiated accusations and erroneous arguments. My feeling is that
SOME readers will be swayed by false logic and should be given a chance
to see that there are other opinions. But many (most?) readers have a
short fuse and just skip reading posts which they find falacious. In the
end the vociferous posters (who are often the ones presenting the
unsubstantiated arguments) have the most perseverance and POTENTIALLY
sway the less experienced readers. I try to offset that.
(snip)
>I'd be very interested to know what people think is a good way of evaluating
>ones skill level against JF, in a reasonable number of games. Surely there
>must be some test that does not involve thousands of games to identify a
>slight edge, or what is the point of all the tournaments?
As Claes has pointed out, it's usually not that simple. The only way
that it takes only a few trials (if you are looking for statistical evidence)
is if the two combatants are VERY SEPARATED in skill level. That is not
likely to be the case when looking for someone (something?) who is better
than JF.
There IS a new method (though not foolproof--what else is new) to compare
one's skill vs. JF WITHOUT having to grind out thousands of games. Record
your match vs. JF and then have Snowie go through it and be a referree. I
would guess that after about a dozen 11-point matches you will have decent
indication of who is better (in the 'eyes' of Snowie, that is). But
I can already hear the counterarguments: "JF and SW are robots. OF course
they are going to side with each other AGAINST the (alien) human species!"
Chuck
bo...@bigbang.astro.indiana.edu
c_ray on FIBS
Murat Kalinyaprak
> Ian Shaw wrote:
>> I think Murat deserves to be cut a bit of slack. It is OK
>> to play BG whilst not knowing anything about statistics,
>> RNGs, Neural Network Development, or even doubling and
>> tournament play.
> True, but I got on his case NOT for these reasons, but for
> breaking my (self-imposed and admittedly subjective) "rules
> of r.g.bg civility":
I'm sorry if my style doesn't please some readers but
I doubt that I'll do anything differently in order to
accommodate particular people's "rules"...
> 1) Profanity is INEXCUSIBLE. (Murat has NOT broken this
> rule, but...)
What is this "but..." addition supposed to mean? I do
occasionally break this rule. If I haven't done so in
this group yet, in the nearly 100 articles I posted
here, I must be doing better than my average. I don't
go as far as using four-letter words, but I may say
"darn, hell, bullshit" etc. For someone who was trying
to be sarcastic about children under the age of 102,
I hope you are old enough to hang around newsgroups
where you risk being exposed to such words...
> 2) Strong accusations (such as of cheating or 'dropping')
> require strong evidence. (See Murat's post of July 8 for
> his intial comments.) See his many other posts for his
> failure to provide the evidnece.
I'm kind of getting a little tired of being "accused
of accusing"... I think I know what you are referring
to, so let's get right down to it. If somebody I don't
know said that he was 7 feet tall, I would take his
word for it. But if somebody said that the world was
flat, or there was a God, or there wasn't a God, etc.
I wouldn't have to take his word for it, even at the
risk/expense of "offending him" or "being myself
accused of accusing him of whatever". I hope the
difference between the two is clear enough to not
need more words.
So, if Fredrik, you or any other JF advocate had said
nothing more than "JF doesn't cheat", I may have been
still willing to just take his/your/their word for it
(*even though* JF is not completely unknown to me).
But, the seed/counter scheme and arguments based on it
(including his own), etc. show that you all have gone
beyond merely expecting people to just take your word
for it, and instead you went on to offer some sorts of
"proofs". (From here on I may just use "you" to address
him/you/them). Thus, people can no longer be expected
to take your word for it either, and it becomes a case
of your "proof" being either good enough or not. If
it's not good enough, then you have no justification
to feel offended, insulted, accused, etc. In other
words, having attempted to offer proof to eliminate
suspicions and falling short of it, in my opionion,
is actually worse than having never tried.
Also, if what you offer as "proof" is not good enough,
then what you are expressing doesn't go beyond being
merely a belief/opinion (i.e. that JF doesn't cheat).
And this applies to Fredrik's arguments also.
You may ask how can that be, since Fredrik is the one
who developed JF...? Let me explain why I believe so.
I had previously expressed that even an unintentional
flaw may rightfully be perceived as "cheating" by the
users (for all practical purposes), if it causes them
any degree of disadvantage at all. As the late bug I
discovered illustrates, Fredrik himself may not be
aware of such flaw/s if they exist beyond his will
and capacity. I'm not saying that this bug is one such
case but just using it as an example to illustrate
what is "possible".
In conclusion, when some unknown amateur newcomer to
this newsgroup expresses unsubstanciated opinions
about JF's cheating, they should not be taken to be
any more than just that. But at the same time, their
unsubstanciated opinions are just as good as yours,
Fredriks or anyone else's unsubstanciated opinions.
Talking down to those newcomers based on non-proofs
is plainly unjust. If it goes to the extent of saying
things like "If you are too dumb to understand this
proof, then (whatever)..." or "If this proof doesn't
relieve your paranoia, go see a shrink...", etc. then
I would consider it as outright insulting to those
people. I got this deep into these discussions mainly
because I myself felt annoyed/insulted/offended/etc.
by the way some people (later including me also) were
put through a certain "treatment" by some writers in
this newsgroup...
It looks like this will end up being quite a long post,
but whoever was interested enough to read this far will
probably find the rest interesting as well. Before we
go on further though, let's clarify a couple of things
here. One: nobody argues that JF cheats by moving pieces
incorrectly, etc. All suspicions expressed seem to focus
on dice rolls. Two: in a debate intended to be based on
logical arguments, arguments make themselves. It doesn't
matter who voices them and they won't go away just by
not voicing them. This is the good thing about logic. It
doesn't need me, you or anybody else. It exists above
and beyond us out there. If we can get an occasional
glimpse of it, then we may consider ourselves lucky...
Now, back to your comments. BTW, I'm not even sure if it
would be worth responding just to you, but I'm doing
it in a way taking you as perhaps representing a certain
kind of people in this group, and my comments intend to
address hopefully a larger audience. This said, I'm also
getting tired of being accused of not substanciating my
arguments.
First of all, I'm not sure if I even need to do that to
begin with, since as long as there is room for me to do
so, I should be able to speculate as I wish. I'm just
analyzing/considering all possibilities and I don't
know why this should strike some people as unusual?,
while they themselves analyze all possible moves for a
given dice roll, etc. all day long.
But let's say that I should feel some responsability to
substanciate them and let's also say that some of my
speculations/arguments may be too wild/vague, etc. for
people to respond to them. But what about the ones that
relate to numbers, averages, ratios, etc. For example,
a few days ago I gave some numbers about the length of
matches I had played against JF, overall average value
of dice rolls in those games, etc. How come I heard no
responses on such concrete, observation based issues I
had raised...?
I don't mind putting my own time and effort into this
and I don't even mind that you guys don't/can't respond.
But I don't think it's right for people to ignore what
I'm writing and keep whining/complaining that I don't
(or even try to) substanciate my arguments.
Related to this issue, one can also ask how can anybody
substanciate anything about something unknown (i.e. kept
secret)? How do you expect me to do this about JF's dice
rolls for example, when I don't have access to neither
the algorithm nor even a convenient means of analyzing
its output? Can you guys do or have ever done what you
are asking from me or others? Can you, for example, tell
me that you have analyzed X million JF dice rolls and
have found no patterns, that it doesn't roll too many
doubles/high numbers or whatever else...? How is anyone
supposed to accomplish what you are asking for? Hack and
reverse-engineer JF's compiled code? If so, then maybe
all that is needed is Fredrik's blessing that it would
be ok to it...
Most of the arguments about cheating have already
turned into general/hypotethical arguments about how
could it be possible for any computer backgammon game
to cheat. I thought my previous article would be the
last one on this subject (for at least a long while,
unless/until something new came about) but being still
harped on, I couldn't withold from responding.
I have previously argued that the stage and precedents
have been set long ago on this issue. At the expense
of repeating myself, I can't understand why Fredrik
hasn't prevented/put an end to suspicions about JF's
dice rolls by just publishing its algorithm.
I have seen may arguments (including his own) trying
to dissociate JF's playing abilities/stregths from its
dice rolls. I accept those arguments as valid and they
effectively trivialize JF's dice rolling function to
the level of being a mere "accessory" as an additional
convenience to users. Anyone can indeed download half
a dozen freeware dice rolling programs from the Internet
in a matter of minutes.
I realize that the seed/counter mechanism serves other
purposes involving re-rolling the same set of dice, but
offering it as a proof of JF's not cheating is actually
counterproductive, while there are simpler and more
effective ways of accomplishing it (like publishing the
algorithm). It's only natural for people to ask: "If
there is nothing to hide, why is something hidden?". Of
course, this argument relies on my not being able to
come up with a *good reason* for not making JF's dice
rolling algorithm public and can be defeated if you or
anyone else can offer a *good reason* for it.
If you or others intend to keep harping on me, I think
it would be fair that you provide some responses on
the issues I'm raising, don't you think...? Otherwise,
repeating the same "bla, bla..." doesn't do much good.
> (No, finding a bug in the RNG input routine is not
> sufficient evidence of cheating!!!!!!)
May I ask how do you know this...? I had already let
it go by wishing him good luck fixing it and was just
going to wait and see when/how it's fixed. So, I'm not
claiming that it's evidence of "cheating" but I'm just
asking you how do you know that it is not...?
Let me also explain why I had asked him for further
details. If I remember correctly, JF was developed
using C++? I don't know C++ but as far as I know,
numbers are commonly stored/interpreted as signed
integers and ranges go like this:
16 bit = -32,768 to +32,767
32 bit = -2,147,483,648 to +2,147,483,647
Normally, unless one wants to make the user wonder
why he is limited to such a specific range like from
1 to 2,147,483,647, valid input ranges are limited to
and indicated by the next largest number that can be
entered using any digits from 0 to 9. So, JF's seed
values being limited to 9,999 or 999,999,999,999 would
make sense to me. It's said that JF produces random
seed values based on the system clock, which seem to
be checked for being less than 32,768 and this makes
sense to me also. Even if JF uses signed integers for
some and unsigned integers for other purposes, I still
couldn't understand how/what 65535 would have to do
with things. Was it rude for me to want to know...?
I couldn't understand what a magnitude check for
<65535 would mean regarding how this would/could be
fixed either. Fredrik said program should have
checked for this but it was overlooked. That means
the right thing to do today or tomorrow will be for
JF to perform this check. And that means the next
revision of JF will limit seed values selectable by
users to be in the range 1 to 9,999...? Am I out of
line for being curious about this? If this newsgroup
wasn't such an "intimidating":) one, we could even
speculate on how this bug will end up being corrected.
Are you brave enough to tell me what's your guess...?
And what about the 5000+ dice rolls JF rolled at
seed=999999999 and another 5000+ at seed=888888888
when I played two sets of 100 games, without any
invalid numbers...? So, dice numbers don't go
completely berzerk once you pass 65535, but work
quite well with even very high seed numbers. Do you
think I was totally unjustified to think that this
might be an indication of a bug not so trivial...?
(Notice that when I expressed Fredric should be given
a fair chance to respond, I had said that he should
also be given *enough time*, simply because I thought
it may take him a while to figure what was causing the
problem. Frankly, I was very surprized, even impressed
that he posted an explanation so promptly (although an
extremely brief one). Imagine how much more impressed
would I have been if he had made a fixed version of JF
player available as quickly also:)...
I can understand that Fredrik may not be able to post
a long article explaining these, because of time
limitation or other reasons. That's why I had asked
one anonymous poster (who had flamed me) if he could
post an algorithm that would "emulate" this trivial
bug. If anybody else is interested, they can take the
same challenge also...
> 3) Insincerity and hypocrisy are out of place here (and
> everywhere else on the face of the earth, for that matter).
> This one can be tough to judge because often people make
> 'tongue in cheak' comments (and scarcasm) which are subject
> to misinterpretation, for example. But feigning modesty
> with an ego that doesn't fit through a canyon is insincerity,
> in my opinion.
I guess it's possible that your analysis of me may
be accurate, but don't expect me to be too eager to
view myself through your eyes, nor to like being
attacked down to my "modesty". Whether by means of
insincerity, hypocrisy, sarcasm, feigned modesty or
ego, I sure will try my best to give it to you just
the way you are asking for it...
> Put 2) and 3) together and you reach threshold in my book
> for getting a harsh reply post from me. (And I'll bet THAT
> scares you!)
> I've been asked why I bother with these kind of posts and
> replies, since it takes time away from seemingly more
> valuable work (like analyzing positions). Similarly, I've
> asked some others why they DON'T reply to unsubstantiated
> accusations and erroneous arguments.
I can't speak for others, but I'm doing this (i.e.
replying to unsubstantiated accusations and erroneous
arguments) on my own will, not because you asked...:)
> My feeling is that SOME readers will be swayed by false logic
> and should be given a chance to see that there are other
> opinions. But many (most?) readers have a short fuse and
> just skip reading posts which they find falacious. In the
> end the vociferous posters (who are often the ones presenting
> the unsubstantiated arguments) have the most perseverance and
> POTENTIALLY sway the less experienced readers. I try to offset
> that.
Have you ever considered whether you may be one
of those "vociferous" posters, who falls short of
being logical...?
Let me quote you yourself:
"Malcolm Davis,... invited two other world class players
"(Mike Senkiewicz and Nick Ballard) to each play 300 games
"at money play... the stakes were $100 per point. Dice
"were rolled manually (by Davis, I believe) so that the
"typical "JF cheats" arguments would be completely irrelevant.
After reading this, the first question that should
come to mind is: "Why did they roll dice manually?".
Was something wrong with JF's rolling the dice...?
Obviously those guys didn't deem JF's dice rolling
*as good as* manual dice rolling. And by your own
mouth, you have proven that *the only way to make the
typical JF cheats arguments completely irrelevant was
to roll dice manually*...!
Around the time you posted this, I also remember some
people had made the argument that no "respectable"
world-class player had accused JF of cheating, etc.
Well, folks, if you can hear a whisper, there you
have it! A clear/indisputable message regarding JF
from some "respectable" world-class players...
Personally, I think this should be considered outright
insulting to Fredrik. Why didn't this make you stomp
your feet, scream and cry that some world-class players
didn't trust JF's rolling the dice? Why was there a
need felt to eliminate "JF cheats" arguments in this
particular case? Surely those were no *lesser players*
who would ever think of blaming their shortcomings on
JF's cheating, were they...? If JF beat them, people
like you would have never said that those world-class
players lost because JF cheated, would you/they? Some
"unrespectable-lesser player-amateur" might have said
such things but then their opinion surely wouldn't be
valued much next to $100/point, would it...?
Here is your challenge of the day: Find and present some
cases where world-class players played against JF for
$100/point, using JF's automatic dice rolling. If you
can do that, then you can come blame me and other lesser
players, amateurs, whoever else for not deeming JF's dice
rolls *as good as* real life and suspecting that it may
be flawed, rigged or whatever. After that, the next step
for you may be to suggest that the above mentioned players
apologize to Fredrik, since their act can very well be
taken as an indirect/implied "accusation"...? But don't
be too *harsh* on them or *scare* them too much... :)
MK
Ian Shaw
Chuck Bower wrote in message <6qkdks$ue1$1...@jetsam.uits.indiana.edu>...
[snip]
>>I'd be very interested to know what people think is a good way of
evaluating
>>ones skill level against JF, in a reasonable number of games. Surely there
>>must be some test that does not involve thousands of games to identify a
>>slight edge, or what is the point of all the tournaments?
>
[snip]
> There IS a new method (though not foolproof--what else is new) to
compare
>one's skill vs. JF WITHOUT having to grind out thousands of games. Record
>your match vs. JF and then have Snowie go through it and be a referree. I
>would guess that after about a dozen 11-point matches you will have decent
>indication of who is better (in the 'eyes' of Snowie, that is). But
>I can already hear the counterarguments: "JF and SW are robots. OF course
>they are going to side with each other AGAINST the (alien) human species!"
>
Twelve 11-point matches does indeed seem "reasonable", just based on gut
feeling. You would not even have to submit to Snowie for analysis; the score
would tell you enough. Snowie would just tell you _where_ you went wrong.
I have a theory though, that it would be possible to achieve better results
against a bot than against a human expert. For example, I might decide to
use Shaw's rule in my 11-point matches, which states:
"If you aren't ABSOLUTELY sure whether the position is a take
or a pass, then it is ALWAYS correct to take, unless the last
pair of rolls (the expert's, then your's) went very badly for
you. I'm not kidding!"
The theory behind this is that an expert is going to make sure that she
doesn't lose her market by doubling when you still have a take. Therefore
you have a take. Q.E.D.
A human expert will quickly realise (from the quality of your checker play,
maybe) that you don't _know_ the correct action and are always taking. She
will then start to double later and thus be able to increase her equity. The
bot, on the other hand, is programmed to double at the optimum time and
won't adjust its play.
Another possibility is that by playing your checkers sub-optimally you will
throw off JFs move evaluations when it is looking ahead (Yes, this sounds
like a Murat Theory (TM) :o). Presumably, when looking ahead two or three
plies, it chooses the move which gives it the greatest equity over the next
21 x 21 X 21 x 21 combinations (two-ply) or 21^6 (three ply) of rolls. In
doing this it *must* assume that you will play the best possible (highest
equity) moves on your rolls. Therefore, the set of positions it is analysing
to help choose its current move may be radically different from what will
actually happen when playing a non-expert. In fact, the deeper it looks, the
more inaccurate the reference space becomes.
It may even be that JF is better at level 5 against a non-expert than at
levels 6 & 7. I am assuming here that at level 5 JF merely evaluates the
possible moves for the current roll; at level six it looks two plies ahead
and at 7 it goes as far as the Time Factor allows. correct me if I'm wrong.
When playing a human, the expert will realise where the other player is weak
and exploit the weakness by tending towards those positions. The bot will
assume that its opponent has no weaknesses and will make the theoretically
correct move.
Note that I am not claiming that a weak player will do better against a bot
than an expert. I AM suggesting that a weak player is likely to achieve
better results agianst a bot than against an expert of similar ability to
the bot.
So much for my theory, I'm off to get thrashed by JF again...
Regards,
Ian
Dan Frank
Subject: Re: What is a fair test of ability?
..
> There IS a new method (though not foolproof--what else is new) to compare
>one's skill vs. JF WITHOUT having to grind out thousands of games. Record
>your match vs. JF and then have Snowie go through it and be a referree. I
>would guess that after about a dozen 11-point matches you will have decent
>indication of who is better (in the 'eyes' of Snowie, that is).
I wonder mightily about the above - not for the message itself, but from whom
it comes: when it comes from an scientist, who deserves that name, it is a
Big (Bang) Blunder! (Blunders happen to anybody - errare ....)
To put Snowie as a referee - although very convenient because of the export/import
capabilities - brings no further light (*), "a fair test of ability" can't be
obtained at all like that.
To put Snowie as a referee, in fact as a Supreme Judge, means to give it credit as
absolutely the best, at least in comparision with both, Murat and JF! The "though
not foolproof" at the begining of the paragraph doesn't change anything: you put
it for that aim as judge or you don't!
(*) see differences between JF and Sw and the postings about the non-absolute-
reliability of neural nets from Brian Sheppard.
Chuck Bower
Ian Shaw <ian....@riverauto.co.uk> wrote:
>Chuck Bower wrote in message <6qkdks$ue1$1...@jetsam.uits.indiana.edu>...
>
>[snip]
>> There IS a new method (though not foolproof--what else is new) to compare
>>one's skill vs. JF WITHOUT having to grind out thousands of games. Record
>>your match vs. JF and then have Snowie go through it and be a referree. I
>>would guess that after about a dozen 11-point matches you will have decent
>>indication of who is better (in the 'eyes' of Snowie, that is).
(snip)
(and Ian's response:)
>Twelve 11-point matches does indeed seem "reasonable", just based on gut
>feeling. You would not even have to submit to Snowie for analysis; the score
>would tell you enough. Snowie would just tell you _where_ you went wrong.
I disagree, and I think you missed an important point (but don't feel
bad because I seem to be doing a poor job today of making my points
understandable). If you just look at the scores from the dozen matches,
you can't really tell if it was predominantly luck or predominantly skill
(or both or neither(?)...) that led to the outcome. Twelve matches of
length 11-points each is insufficient to make a strong statistical conclusion
(unless maybe if the score is 12-0 or 11-1, which will not be the case with
most seasoned players). However, Snowie ignores the luck and just says how
you did with the dice given you. (Actually there still is some luck present.
One player may be given "easy" decisions and the other player "difficult"
decisions. But this method goes a long way towards eliminating the luck.)
(Ian continued:)
>I have a theory though, that it would be possible to achieve better results
>against a bot than against a human expert. For example, I might decide to
>use Shaw's rule in my 11-point matches, which states:
>
> "If you aren't ABSOLUTELY sure whether the position is a take
> or a pass, then it is ALWAYS correct to take, unless the last
> pair of rolls (the expert's, then your's) went very badly for
> you. I'm not kidding!"
>
>The theory behind this is that an expert is going to make sure that she
>doesn't lose her market by doubling when you still have a take. Therefore
>you have a take. Q.E.D.
Interesting point with some validity, IMO. But, experts make mistakes
(although you can argue "less likely than I!")
>A human expert will quickly realise (from the quality of your checker play,
>maybe) that you don't _know_ the correct action and are always taking. She
>will then start to double later and thus be able to increase her equity. The
>bot, on the other hand, is programmed to double at the optimum time and
>won't adjust its play.
Who is this expert? How does she figure out "Ian's rule" so quickly?
I think you give her too much credit. However, it is true that bots that
make their play decisions "cubeless" (and "scoreless" in the case of a match)
will sometimes err. The classic example is when there is a killer roll that
will allow a cash next time with play A and cash MOST of the time with
play B, but play B comes out higher in cubeless equity. The bot (playing
"cubeless") chooses play B, but the "expert" human correctly chooses A.
>Another possibility is that by playing your checkers sub-optimally you will
>throw off JFs move evaluations when it is looking ahead (Yes, this sounds
>like a Murat Theory (TM) :o). Presumably, when looking ahead two or three
>plies, it chooses the move which gives it the greatest equity over the next
>21 x 21 X 21 x 21 combinations (two-ply) or 21^6 (three ply) of rolls. In
>doing this it *must* assume that you will play the best possible (highest
>equity) moves on your rolls. Therefore, the set of positions it is analysing
>to help choose its current move may be radically different from what will
>actually happen when playing a non-expert. In fact, the deeper it looks, the
>more inaccurate the reference space becomes.
Well there is one thing I do agree with you in this parargraph--it
does appear Murat-like. However, I don't think it holds up. If you consider
backgammon as Danny Kleinman's "(American) football field" it looks like you
are saying it could be better to lose yardage (at least that is how I
consider your term "play sub-optimally"). You make a worse play (in the
eyes of JF). To me that is like saying "I think I'll get tackled for a
loss here so that the defense will have more trouble defending 3rd and 20
yards to go as compared to defending 3rd and 10 yards to go." It's still
the same (strong) defense. Now you have to move the ball even further than
if you had just made the "optimal" play.
(snip)
>When playing a human, the expert will realise where the other player is weak
>and exploit the weakness by tending towards those positions. The bot will
>assume that its opponent has no weaknesses and will make the theoretically
>correct move.
True. That is one reason why a "collaboration" of strong human and strong
bot should be even stronger than either alone. Doubles, anyone? ;) But how
much does this change the equity (for a given play) and how often does the
expert human get a chance to try it? (I'm not in a good position to answer.)
Also, how many games (or matches) does it take the human expert to find
the weaknesses? And how often does s/he think s/he finds a weakness and
actually be wrong? Not simple questions, but all of them are relevant.
>Note that I am not claiming that a weak player will do better against a bot
>than an expert. I AM suggesting that a weak player is likely to achieve
>better results agianst a bot than against an expert of similar ability to
>the bot.
True, but besides one recent boaster poster, who do you know who is of
"similar ability?"
Chuck Bower
Dan Frank <10001...@CompuServe.COM> wrote:
>bo...@bigbang.astro.indiana.edu (Chuck Bower) wrote:
> Subject: Re: What is a fair test of ability?
>
>..
>> There IS a new method (though not foolproof--what else is new) to compare
>>one's skill vs. JF WITHOUT having to grind out thousands of games. Record
>>your match vs. JF and then have Snowie go through it and be a referree. I
>>would guess that after about a dozen 11-point matches you will have decent
>>indication of who is better (in the 'eyes' of Snowie, that is).
(and Dan Frank responds:)
>I wonder mightily about the above - not for the message itself, but from whom
>it comes: when it comes from an scientist, who deserves that name, it is a
>Big (Bang) Blunder! (Blunders happen to anybody - errare ....)
Well, I'll give Dan credit for making a clever "tongue in cheek"
comment, but now it is time to say something serious. I only have one
computer account with internet access. That is my work account and the
headings of my posts and e-mails are canned there. I believe that titles
and reputations should take a VERY DISTANT back seat to the substance of
a post. Just because I have a title which some might consider intellectual
doesn't mean I know beans about what I say. There are a LOT of ignorant
scientists, and I may be one of them. Read the content. If it is
garbage, then file it there. I guarantee that I will do the same with
ALL the posts I read.
You may have seen my rants about 'authoritarianism'. "So-and-so is
an authority, so s/he is likely to be right." BALONEY! This is just an
excuse for turning off the brain. Read the logic (or lack thereof) and
make a decision regardless of the name and/or title attached. (Now, please
excuse me while I descend from my soapbox, hoping not to stumble and land
on my face.)
Donald Kahn
Bower) wrote:
>In article <ONpAswK...@nih2naaa.prod2.compuserve.com>,
>Dan Frank <10001...@CompuServe.COM> wrote:
>
>>bo...@bigbang.astro.indiana.edu (Chuck Bower) wrote:
>> Subject: Re: What is a fair test of ability?
>>
>>..
>>> There IS a new method (though not foolproof--what else is new) to compare
>>>one's skill vs. JF WITHOUT having to grind out thousands of games. Record
>>>your match vs. JF and then have Snowie go through it and be a referree. I
>>>would guess that after about a dozen 11-point matches you will have decent
>>>indication of who is better (in the 'eyes' of Snowie, that is).
>
> (and Dan Frank responds:)
>>I wonder mightily about the above - not for the message itself, but from whom
>>it comes: when it comes from an scientist, who deserves that name, it is a
>>Big (Bang) Blunder! (Blunders happen to anybody - errare ....)
>
> Well, I'll give Dan credit for making a clever "tongue in cheek"
>comment,...
I don't understand his comment, but like the rest of his comments, it
has little or no value, because of his lack of any expert, or even
intermediate, knowledge concerning the game.
Eleven 11-point matches matches, merely looking at the results would
be more than good enough for me to decide whether I should consider
playing the individual for a meaningful stake, by which I mean more
than I can afford to lose.
If you assign a skill differential between the players as to a single
11-point match, (for instance, A has a 55% chance of winning), you can
derive a strict algebraic expression for the chance of A winning at
least 6 of them. It must be around 75%.
IMO, letting Snowie analyze it is even MORE convincing.
deekay
bshe...@hasbro.com
don...@alpin.or.at (Donald Kahn) wrote:
> On 11 Aug 1998 00:16:30 GMT, bo...@bigbang.astro.indiana.edu (Chuck
> Bower) wrote:
>
> >In article <ONpAswK...@nih2naaa.prod2.compuserve.com>,
> >Dan Frank <10001...@CompuServe.COM> wrote:
> >>bo...@bigbang.astro.indiana.edu (Chuck Bower) wrote:
>
> >>it comes: when it comes from an scientist, who deserves that name, it is a
> >>Big (Bang) Blunder! (Blunders happen to anybody - errare ....)
For the record, I regard Chuck Bower to be an exceptionally trustworthy source
of insight into backgammon. A scientist to the core.
> Eleven 11-point matches matches, merely looking at the results would
> be more than good enough for me to decide whether I should consider
> playing the individual for a meaningful stake, by which I mean more
> than I can afford to lose.
>
> If you assign a skill differential between the players as to a single
> 11-point match, (for instance, A has a 55% chance of winning), you can
> derive a strict algebraic expression for the chance of A winning at
> least 6 of them. It must be around 75%.
Watch out, DK! The weaker player's chance of winning a short series is less
than you think.
My calculation shows that the chance of winning an 11-match series (i.e., the
chance of being the first to win 6 matches) when your chance of winning a
single match is 55% is only 63.3123%. If this seems too low, keep in mind that
the stronger player's average outcome in an 11-match series is just a shade
over 6-5, so if the weaker player gets just a little lucky then he can win it
all.
You are not alone, DK. Most people believe that short matches have far
greater predictive significance than they really have. (That's a good thing
for professional sports. If the general public consisted only of
statisticians then the NHL playoffs would still be going on.)
> IMO, letting Snowie analyze it is even MORE convincing.
For my money, this is the acid test. Every game consists of dozens of plays,
so over the course of a match you can measure the quality of hundreds of
playing decisions. Foe each decision you obtain a continuously valued error
estimate which measures the best you could have accomplished with the dice.
This method works so well that I use it to analyze the quality of my neural
networks.
One hundred games are sufficient to evaluate a player's checkerplay. To judge
cube handling you need more games than that.
Warm Regards,
Brian Sheppard
Ian Shaw
bshe...@hasbro.com wrote in message <6qpfta$ov$1...@nnrp1.dejanews.com>...
[snip]
>
>My calculation shows that the chance of winning an 11-match series (i.e.,
the
>chance of being the first to win 6 matches) when your chance of winning a
>single match is 55% is only 63.3123%. If this seems too low, keep in mind
that
>the stronger player's average outcome in an 11-match series is just a shade
>over 6-5, so if the weaker player gets just a little lucky then he can win
it
>all.
>
Excuse me if I'm being dim. Please would you clarify whether you are
calulating:
a) the probablity of winning six out of eleven eleven-point matches, given
that the probablity of wining a single GAME is 55%.
b) the probablity of winning six out of eleven eleven-point matches, given
that the probablity of wining a single MATCH is 55%.
If we wanted, we presumably could build a complete match equity table using
Tom Keith's method for different relative strengths.
e.g.
P(Strong Player's Gammon) = 12%
P(Strong Player's Win) = 45%
P(Weak Player's Win) = 37%
P(Weak Player's Gammon) = 6%
Regards
Ian Shaw
Ian Shaw
Chuck Bower wrote in message <6qo163$uj7$1...@flotsam.uits.indiana.edu>...
>Ian Shaw <ian....@riverauto.co.uk> wrote:
[snip]
I'm not suggesting that you deliberatly play less well; the weaker player
makes the best move as he sees it. Let's see if I can explain better.
What I believe is that when looking ahead n-plies, a computer program does
not consider WEAK future responses when working out what to do NOW.
I am supposing that when a computer program is trying to determine its best
move by looking ahead 1 ply, it:
a) computes all the possible positions arising from the current position
plus three rolls (Computer0, Human1, Computer1)
b) evaluates the equity of all these positions
c) discards all positions which give the Computer less than the best
equity for each of its 21 possible rolls at Computer1, for each possible
position after Human1.
d) ASSUMES that the Human will make the highest equity move for each of
her possibilities at Human1, and discards the others.
e) selects the move for Computer0 which will give the human the lowest
average equity by the end of Computer1.
I.E. the computer seeks to reduce the human's equity after Computer1 by
selecting a move for Computer0 based on the best possible play by both
sides.
If I could draw this it would look like a tree, pruned to have 21 branches
going off from the trunk, and twenty one twigs going off from each branch.
Each twig has an equity value, and the overall equity is the average of all
these (weighted for the probabilities of rolling doubles etc).
The critical step is (d). If the human does not make the best move, then the
set of positions it is weighing up for Computer1 is not going to occur. This
makes the assessment of the best move for position Computer0 (step e)
potentially incorrect.
The look-ahead method only works well when the opponnet makes the expected
moves.
Against a lesser opponent the computer would achieve better results by not
looking ahead, and simply playing to achieve the highest equity position
after Computer0. Hence my original theory that JF level 5 may have better
results against a weak player than level 6 or 7.
OK, the program could be refined to evaluate the top two or three human
moves, give them weightings for the probability for each (i.e. the
probability of a mistake), and proceed on that basis. Estimating the
probability of a mistake would be very uncertain; is the computer going to
look up your FIBS rating? Therefore I suspect that this is not done.
By contrast, an expert may think, "I know remember that play Y is
theoretically best. However, Ian has a poor grasp of timing considerations;
let's make play X which tends to get us into a prime vs prime game where
he'll make mistakes."
Does this make any sense?
Regards,
Ian
Phill Skelton
>
> bshe...@hasbro.com wrote in message <6qpfta$ov$1...@nnrp1.dejanews.com>...
> [snip]
> >
> > My calculation shows that the chance of winning an 11-match series
> > (i.e., the chance of being the first to win 6 matches) when your
> > chance of winning a single match is 55% is only 63.3123%. If this
> > seems too low, keep in mind that the stronger player's average
> > outcome in an 11-match series is just a shade over 6-5, so if the
> > weaker player gets just a little lucky then he can win it all.
> >
>
> Excuse me if I'm being dim. Please would you clarify whether you are
> calulating:
>
> a) the probablity of winning six out of eleven eleven-point matches,
> given that the probablity of wining a single GAME is 55%.
>
> b) the probablity of winning six out of eleven eleven-point matches,
> given that the probablity of wining a single MATCH is 55%.
It is option b), though the length of each match is not specified -
it's just 11 matches of indeterminate length, though the same
length in each case to keep the probability the same. If it were
11, 11-point matches, with 55% chance of winning each GAME, then
the better player would win the series over 70% of the time.
> If we wanted, we presumably could build a complete match equity
> table using Tom Keith's method for different relative strengths.
The mec program (available at Steven Turner's page IIRC) produces
match equity tables for any length match, depending on the relative
strengths and gammon rate (though the fraction of each player's wins
that are gammons is the same for both players).
Phill
Phill Skelton
> I am supposing that when a computer program is trying to determine
> its best move by looking ahead 1 ply, it:
> a) computes all the possible positions arising from the current
> position plus three rolls (Computer0, Human1, Computer1)
> b) evaluates the equity of all these positions
> c) discards all positions which give the Computer less than the
> best equity for each of its 21 possible rolls at Computer1,
> for each possible position after Human1.
> d) ASSUMES that the Human will make the highest equity move for
> each of her possibilities at Human1, and discards the others.
> e) selects the move for Computer0 which will give the human the
> lowest average equity by the end of Computer1.
>
> I.E. the computer seeks to reduce the human's equity after Computer1
> by selecting a move for Computer0 based on the best possible play by
> both
> sides.
>
> The critical step is (d). If the human does not make the best move,
> then the set of positions it is weighing up for Computer1 is not
> going to occur. This makes the assessment of the best move for
> position Computer0 (step e) potentially incorrect. The look-ahead
> method only works well when the opponnet makes the expected moves.
> Against a lesser opponent the computer would achieve better results
> by not looking ahead, and simply playing to achieve the highest
> equity position after Computer0. Hence my original theory that JF
> level 5 may have better results against a weak player than level
> 6 or 7.
I would expect that a computer looking ahead 3 or 5 ply would still
play better than just looking at only it's possible moves, simply
because it will spot more situations where bad rolls would force it
to do something horrible (unless the humans play was truly awful).
What it COULDN'T do, that a human expert could, is optimise it's
equity gain per roll for a given opponent. If the human player always
plays the computers 2nd choice move (for example), then a computer
that knew this would do better against the human than a computer
that always assumed it's opponent would make the best possible play
(Read that again and it might make more sense), because it would
be optimising it's equity gain from the human's mistakes. However,
I expect that both these multiple ply lookahead 'bots would do better
than a computer that only looked at 1 move.
Phill
Chuck Bower
Ian Shaw <ian....@riverauto.co.uk> wrote:
(snip)
>I'm not suggesting that you deliberatly play less well; the weaker player
>makes the best move as he sees it. Let's see if I can explain better.
>
>What I believe is that when looking ahead n-plies, a computer program does
>not consider WEAK future responses when working out what to do NOW.
(snip)
I think the computer assumes that its opp (and itself, for that matter)
will make the "best" play at each turn.
>The look-ahead method only works well when the opponnet makes the expected
>moves.
I don't know your definition of "well", but I disagree (maybe just
semantically). It may not make the best play for the given opponent,
but it still makes a strong play.
>Against a lesser opponent the computer would achieve better results by not
>looking ahead, and simply playing to achieve the highest equity position
>after Computer0. Hence my original theory that JF level 5 may have better
>results against a weak player than level 6 or 7.
I don't agree with this either. Just because it doesn't pick the
absolute best play at level-7 to take advantage of its opponent doesn't
mean it does choose the best play at level-5. It COULD be the case but
it doesn't seem that way to me.
(snip)
>By contrast, an expert may think, "I know remember that play Y is
>theoretically best. However, Ian has a poor grasp of timing considerations;
>let's make play X which tends to get us into a prime vs prime game where
>he'll make mistakes."
OK, here I agree. A human can adjust his/her play to take maximum
advantage of his/her opponent's weaknesses (and thus play differently than
against another expert). This certainly works sometimes. I actually thought
you meant that when a robot is playing a sub-expert that the sub-expert
could do better by making what JF considered a "non-best" play. THAT
I disagreed with (and still do) but apparently that wasn't your point...
bshe...@hasbro.com
!remove!this!ph...@sun.leeds.ac.uk (Phill Skelton) wrote:
>
> I would expect that a computer looking ahead 3 or 5 ply would still
> play better than just looking at only it's possible moves, simply
> because it will spot more situations where bad rolls would force it
> to do something horrible (unless the humans play was truly awful).
I have played thousands of games against JF on levels 5, 6, and 7. I keep
careful records of every game outcome. My statistics show that my results do
not depend on which level JF is playing at. (Of course, my play could be
"truly awful." :-))
The effect of searching is to make JF Level 7's play more resilient against
JF Level 5, through precisely the mechanism that Ian Shaw described. When JF
Level 7 confronts me, it is playing a very different opponent than when it
confronts JF Level 5. Let's look at specifics.
On the good side, my play is stronger than JF's in certain specific
situations. For the most part these situations involve timing or cube
handling. I gain an advantage over JF Level 6 and 7 to the extent that its
search engine pursues weaker variations than I would.
On the downside, my play is weaker than JF's in other situations. For the most
part these situations involve checkerplays in double-edged positions. To the
extent that JF Level 5 creates such positions (even perhaps when it is
theoretically incorrect to do so!) then it gains an advantage over me compared
to JF Level 7, which plays situations more "tightly."
As I mentioned before, my overall results against various levels of JF do not
differ significantly. (In fact, the difference is less than 1% after thousands
of games.) My conclusion is that searching in backgammon is overrated.
The effect of searching, in general, varies according to the quality of two
underlying models. One model is the evaluation model. The stronger the
evaluation function is, the less the gain from searching. Another model is the
search-space model. The more faithfully the search-space model corresponds to
reality, the better the search results are.
In backgammon we have neural networks that epitomize accuracy in evaluation.
The mean error of a JF evaluation is approximately 3.5%. (Specifically, for
the statisticians in the audience, the root-mean square of the sum of the
squared deviations of *all* of JF's network outputs is about 3.5%. This
implies that the average accuracy in a single output (like gammon chances) is
better than that.) It follows that the gain from search should be small.
And as Ian (and I in previous posts) have pointed out, the search-space models
of backgammon programs do not correspond to reality. There are two specific
defects.
First is that the search model is cubeless, whereas the game is cube-using.
This introduces huge inaccuracies into the search! If you evaluate a terminal
node of a search tree as having an equity of 0.700, then the cubeless search
engine is making an evaluation error of 0.300 on that node, because the real
situation is a double/drop. Think of all the time wasted on making the neural
network super-accurate, when the search engine introduces errors of ten times
that magnitude!
Second is the selective-search model. The assumption is that it is OK to use
the neural network to choose moves within the search tree. This introduces
some error because the neural network is occasionally wrong. It introduces
additional error because the neural network might not be representative of
its opponent's play.
These problems are structural, and require additional programming effort.
Cube- using search engines are more complicated than cubeless. Full-width
search is more complicated (and time-consuming) than selective search.
Validating a model of the opponent's play is difficult. But I maintain that
programmers must address these difficult issues if they are to prove their
assertion that programs surpass the best humans.
Warm Regards,
Brian
Matthias Nilsson
I need a backgammon-playing shareware program for win95
Which is your favourite and where can I download it?
Matthias
Ian Shaw
Regards
Ian
Matthias Nilsson wrote in message <6qslhn$r$1...@cubacola.tninet.se>...
Bob Kaplan
I have a couple of questions about JellyFish which does not have a help
file:
1. Is there a way to save the settings: Your name, 3 point match, etc.
2. Is there a way to keep a cumulative score over time?
Thanks a lot,
bob
Ian Shaw wrote in message <6quico$a...@news3.force9.net>...
Murat Kalinyaprak
>Murat Kalinyaprak wrote:
>>For being someone who can block 5 character
>>long lines, wouldn't you wish your name was
>>something like James, Murat or Steve...? :)
>Marina For another Then, if you had
>is not column, you your Thesaurus
>so bad can provide to hand, you
>either a different could do a
>if you width for a tasteful
>can be longer name shaped
>blithe - fibs name bit.
>with 6 or possibly
>letter a different
>words. server sig.
>Marina mas on fibs
I have I suppose I Glad
to say am lucky to to see
you're have a last that you
indeed name that's joined the
a very long enough club also. I
crafty to fit this wonder if many
one at width of 11 more will end up
these. characters. with this disease?
MK
Murat Kalinyaprak
>In <6qmvut$f...@news3.force9.net> Ian Shaw wrote:
>>Another possibility is that by playing your checkers
>>sub-optimally you will throw off JFs move evaluations
>>when it is looking ahead (Yes, this sounds like a
>>Murat Theory (TM) :o).
Ian, I like the joke but I hope it's wasn't sarcastic.
Since I'm not allowed to "feign modesty" anynore, I was
going to give you the credit of being one of the very
few people in this group who understand the complexities
of my superior mind... :) Don't ruin the opportunity...
>> Presumably, when looking ahead two or three plies, it
>> chooses the move which gives it the greatest equity
>> over the next 21 x 21 X 21 x 21 combinations (two-ply)
>> or 21^6 (three ply) of rolls. In doing this it *must*
>> assume that you will play the best possible (highest
>> equity) moves on your rolls. Therefore, the set of
>> positions it is analysing to help choose its current
>> move may be radically different from what will actually
>> happen when playing a non-expert. In fact, the deeper
>> it looks, the more inaccurate the reference space becomes.
This is all fine but let me make a minor clarification.
The key is that how JF will play is already known (at
least to a great extent) because it will always do the
"best moves". So, the human opponent can't be just a
beginner but must be good enough to at least come close
to knowing what would be statistically considered as
"best moves" in various situations. We are not talking
about inferior moves made "haphazardly" here, but moves
that are made "knowingly". This is important...
Often times the difference between the "best move" and
the "next best move" (or even the "third/fourth/etc.
best move" is statistically minute. Yet, even if the
difference is only 1%, we know that a bot will prefer
it consistently. This is fine and over X million games,
that 1% may be visible/significant and eventually may
pay off. But it may not make much difference in 100 or
even 1000 games. Basing one's play on statistics alone
is a strategy in itself indeed, but a rather poor/sterile
strategy for humans, in my opinion, because our lifespans
are not long enough to play such large numbers of games
and considerably benefit from such minute difference. I
think this would hold true even if we disregarded any
other considerations besides seeing backgammon as a pure
means of gambling.
The question here is whether a human can be creative
enough to quickly devise flexible strategies to extract
any advantage from differences between 1% and 2%, etc.
in the short term.
Another speculation I had made (similar to this one but
not the same) was that moves that would be considered
plain "errors/blunders/etc." could possibly work against
JF also, but that was based on the possibility that there
may be recognizable, predictable patterns in JF's dice
rolls (which we don't know for sure if there is or not).
I thought I would clarify this just in case there is any
confusion, because this one was a whole different subject.
> Well there is one thing I do agree with you in this
> parargraph--it does appear Murat-like. However, I
> don't think it holds up.
Chuck, do you think you could just forget about that this
was proposed by someone named "Murat" and try to give it
a sincere consideration as though it was an idea from
someone named "John"? Give it a try, will you? Better yet,
try it in actual plays against JF or Snowie or some other
bot and see what happens. I'm not suggesting this because
it would mean that if you couldn't, nobody else could do
it either. But because there is a chance that you may be
able to make it work for yourself and become convinced.
>... To me that is like saying "I think I'll get tackled
> for a loss here so that the defense will have more
> trouble defending 3rd and 20 yards to go as compared to
> defending 3rd and 10 yards to go."...
What kind of an anology is this? I don't know much about
American football but surely they don't roll dice to
determine which men will move in what direction and slam
into which men of the opposing team, etc. do they? If
they did that, the above argument could/would very well
apply to football games also.
>... Also, how many games (or matches) does it take the
> human expert to find the weaknesses? And how often does
> s/he think s/he finds a weakness and actually be wrong?
> Not simple questions, but all of them are relevant.
At last, I can agree with some of your comments. In a
previous article, I had myself questioned whether what
I was doing wasn't hurting me more than doing me good.
(In which case I could possibly win by a better margin
by not doing this. But I just can't bring myself to
believe that I could beat JF playing the JF way.)
One experiment I still need to do is to play 100 games
using manual dice to compare the results. I haven't
done it yet with the excuse that I haven't had time,
but the real reason may be that I'm affraid of the
possibility of loosing by a significantly different
margin. Now, don't mistake this as my being affraid
because I would feel bad about having lost against JF,
etc. It's because if that were to happen, then my
suspicions about patterns, etc. in JF's dice rolls
would become even stronger and it could give some
credibility (at least in my eyes) to my opinion that
I was able to beat JF by making tricky moves based on
"hocus-pocus" like predicting its next dice roll, etc.
Given that in the eyes of some readers like you I may
have already created the image of "bad Murat vs. poor
little good JF", wouldn't this be a terrible thing to
happen? Besides, loosing on purpose is much easier than
winning. How would I convince myself that I didn't
subconscientiously withold from trying my best and lost
on purpose, let alone how could I convince others of it?
And just in case you are interested in making more
analytical comments about my behavior or personality,
this is not to "feign honesty"...
>>Note that I am not claiming that a weak player will do
>>better against a bot than an expert. I AM suggesting
>>that a weak player is likely to achieve better results
>>agianst a bot than against an expert of similar ability
>>to the bot.
> True, but besides one recent boaster poster, who do
> you know who is of "similar ability?"
I have no idea who that "boaster" may be :), but now
that I'm not allowed to "feign modesty", I guess I
have to consider myself (at least based on the results
of a couple of hundred games) to be of similar ability
to a certain bot. The consequences of being called an
"expert" may be "unpleasant", but I'll just have to
put up with it... :)
MK
David Montgomery
>Let me quote [Chuck Bower]:
>
> "Malcolm Davis,... invited two other world class players
> "(Mike Senkiewicz and Nick Ballard) to each play 300 games
> "typical "JF cheats" arguments would be completely irrelevant.
>
>After reading this, the first question that should
>come to mind is: "Why did they roll dice manually?".
>Was something wrong with JF's rolling the dice...?
[...]
>Here is your challenge of the day: Find and present some
>cases where world-class players played against JF for
>$100/point, using JF's automatic dice rolling.
Without commenting on Murat's many other points, this is
very easy to understand. Jellyfish's rolls are determined
by the seed and counter. If we have Jellyfish determine
the rolls, then the seed and counter could be set to
produce rolls that favor one side or the other. Who gets
to determine the seed and counter, and how can this person
be trusted by both sides?
If you let JF pick the seed based on the system clock, you
still have problems. If the seed and counter are not noted,
then the side operating the computer could change the seed
or counter to a more favorable setting. If the seed and
counter were noted, there is a tiny possibility that the
side opposing the computer might have some fore-knowledge
of the coming rolls. In any event, a tedious verification
process would be necessary after a game was played to ensure
that JF generated the rolls expected with that seed and counter.
In short, because JF's dice rolls are demonstrably completely
deterministic, they are inappropriate for gambling. For
a user playing against the computer for fun and education, with
no motivation to cheat, JF's rolls are fine.
Also, even if JF's dice roller is completely fair, the program
could be modified by a dedicated cheater so that the program would
cheat for one side or the other. With large amounts of money
on the line, you want to guard against this possibility.
David Montgomery
mo...@cs.umd.edu
monty on FIBS
Bols...@aol.com
defaults to the registered name which in this case is JellyFish. If you
forget to change player two's name and you beat JellyFish, it still says
"JellyFish wins X points" :-)
Bob Olson
norseman on fibs
In article <6qurj2$fdc$1...@nntp2.ba.best.com>,
"Bob Kaplan" <kap...@usernomics.com> wrote:
> Hello Matthias - Folks,
>
> I have a couple of questions about JellyFish which does not have a help
> file:
>
> 1. Is there a way to save the settings: Your name, 3 point match, etc.
>
> 2. Is there a way to keep a cumulative score over time?
>
> Thanks a lot,
>
> bob
>
-----== Posted via Deja News, The Leader in Internet Discussion ==-----
Murat Kalinyaprak
>Murat Kalinyaprak wrote:
>Marina mas on fibs
Murat Kalinyaprak
Murat Kalinyaprak
>In <6qmvut$f...@news3.force9.net> Ian Shaw wrote:
>>Another possibility is that by playing your checkers
>>sub-optimally you will throw off JFs move evaluations
>>when it is looking ahead (Yes, this sounds like a
>>Murat Theory (TM) :o).
Ian, I like the joke but I hope it's wasn't sarcastic.
Since I'm not allowed to "feign modesty" anynore, I was
going to give you the credit of being one of the very
few people in this group who understand the complexities
of my superior mind... :) Don't ruin the opportunity...
>> Presumably, when looking ahead two or three plies, it
>> chooses the move which gives it the greatest equity
>> over the next 21 x 21 X 21 x 21 combinations (two-ply)
>> or 21^6 (three ply) of rolls. In doing this it *must*
>> assume that you will play the best possible (highest
>> equity) moves on your rolls. Therefore, the set of
>> positions it is analysing to help choose its current
>> move may be radically different from what will actually
>> happen when playing a non-expert. In fact, the deeper
>> it looks, the more inaccurate the reference space becomes.
This is all fine but let me make a minor clarification.
confusion, because this one was a whole different subject.
> Well there is one thing I do agree with you in this
> parargraph--it does appear Murat-like. However, I
> don't think it holds up.
Chuck, do you think you could just forget about that this
was proposed by someone named "Murat" and try to give it
a sincere consideration as though it was an idea from
someone named "John"? Give it a try, will you? Better yet,
try it in actual plays against JF or Snowie or some other
bot and see what happens. I'm not suggesting this because
it would mean that if you couldn't, nobody else could do
it either. But because there is a chance that you may be
able to make it work for yourself and become convinced.
>... To me that is like saying "I think I'll get tackled
> for a loss here so that the defense will have more
> trouble defending 3rd and 20 yards to go as compared to
> defending 3rd and 10 yards to go."...
What kind of an anology is this? I don't know much about
American football but surely they don't roll dice to
determine which men will move in what direction and slam
into which men of the opposing team, etc. do they? If
they did that, the above argument could/would very well
apply to football games also.
>... Also, how many games (or matches) does it take the
> human expert to find the weaknesses? And how often does
> s/he think s/he finds a weakness and actually be wrong?
> Not simple questions, but all of them are relevant.
At last, I can agree with some of your comments. In a
>>Note that I am not claiming that a weak player will do
>>better against a bot than an expert. I AM suggesting
>>that a weak player is likely to achieve better results
>>agianst a bot than against an expert of similar ability
>>to the bot.
> True, but besides one recent boaster poster, who do
> you know who is of "similar ability?"
I have no idea who that "boaster" may be :), but now
Gary Wong
> Often times the difference between the "best move" and
> the "next best move" (or even the "third/fourth/etc.
> best move" is statistically minute. Yet, even if the
> difference is only 1%, we know that a bot will prefer
> it consistently. This is fine and over X million games,
> that 1% may be visible/significant and eventually may
> pay off. But it may not make much difference in 100 or
> even 1000 games.
I disagree. By "1%", I assume you mean 1% of the cubeless probability
of winning. If you consistently make an error of this magnitude just
once in every game, then you would expect to lose about 4 points per
100 games (because the average points won per game is about 2), and
about 40 points per 1000 games. Not a huge difference, but it is
significant, and if you make more than one error in a game then they
will add up quickly.
It's not immediately obvious what an error of 1% is, so here's a
concrete example. Suppose you and I both have closed boards; I have
one chequer on the bar and you have 3 chequers scattered around the
outfield. Positions similar to this occur fairly frequently, so I'm
sure any intermediate player has come across them before. If you
could cube it would be a double/drop, so let's assume I own the cube
at 2. Assume you have just rolled some small number like 42. Now, a
novice player playing on your side might decide that since his
opponent is on the bar and not going to get to roll next turn whatever
happens, it really doesn't matter which of the outfield chequers he
moves. However, a stronger player would realise that (this is an
oversimplification to keep the example short) she should ensure she
leaves an even number of crossovers, so that she isn't forced to
expose a blot if she rolls 66. What is the difference in equity
between the even and the odd crossover play? Well, if you leave a
blot and it's hit, you'll be cubed out immediately. You're forced to
leave a blot with the incorrect play on a single roll, 1/36 -- let's
call it 3% to be simple. Your opponent then has 11 shots from the bar
-- let's call it 1/3 to be simple, so overall you are leaving a 1%
chance of losing the game outright with the weak play as opposed to
the safe play. In other words, a difference of 1% CPW.
Now, Murat -- what would you do in this position? If you would be careful
to select the safe play, then I claim that you _are_ sensitive to errors
of 1%. Don't be too critical of others arguing over small differences,
when you do the same yourself.
Cheers,
Gary.
--
Gary Wong, Department of Computer Science, University of Arizona
ga...@cs.arizona.edu http://www.cs.arizona.edu/~gary/
Murat Kalinyaprak
>I'm not suggesting that you deliberatly play less well;
>the weaker player makes the best move as he sees it.
For a while I had thought that we were on the same
wavelength, but this shows that we are not. This
could effectively explain how a "lesser player" may
end up beating JF while "superior players" can't do
it... :) I could understand/expect easily why such
a "therory" might come from some avid JF defenders,
but not why it came from you.
Although I would say that it's possible to "throw
off" JF by making lesser moves while rolling dice
manually, I would not put much into it (not to the
extent of trying it regularly myself anyway).
Let's remember that my initial speculation was based
on the possibility that dice sequences that lead to
difficult and/or impossible positions (i.e. you can
play 3, 4 different ways but get hit no matter how
you play) more often than in real life. I had gotten
this impression because quite often such positions
semed to occur in consecutive or almost consecutive
rolls (sometimes up to half a dozen in a row).
My speculation was that one got cought in these by
making the expected (i.e. good/best) moves and that
if one could see it coming, he could prevent these
sequences/patterns by making an unexpected (i.e.
lesser/wrong) move. Along the way, I had stressed
that such observations could be just my imagination
or that I was noticing them becasue I was looking
specificly for such things. Nevertheless, they were
not speculations completely made in the dark (i.e.
about some computer program that I had never seen
or played against). I had posted some examples of
such situations where even one preemptive move is
not enough but 2-3 in a row may actually save you.
I have the impression that within a span of a few
hours, dozens of examples can be found for this.
The question then may be whether this happens as
frequently in "real life" or even against JF with
manual dice.
Of course, JF itself gets caught in such loops too.
Maybe I should have/still should explore further if
JF gets caught in them as often and how it either
avoids or gets out of them. This could strengthen
or dispel my speculation.
>d) ASSUMES that the Human will make the highest equity
>move for each of her possibilities at Human1, and
>discards the others......
>The critical step is (d). If the human does not make the
>best move, then the set of positions it is weighing up
>for Computer1 is not going to occur. This makes the
>assessment of the best move for position Computer0
>(step e) potentially incorrect.
This is certainly possible but I don't think that
any benefit can be derived from it intentionally.
Even if it had to be an "imaginary" one, one needs
a basis for making such moves. And something very
important also is that any such moves need to be
followed by "correct" moves, in order to justify
the concept of "throwing off". You can't keep
"throwing off" a bot in every move, forever. The
human player must be at least good enough to make
the subsequent moves work or at least be able to
survive the worse consequences of such moves on a
consistent enough basis.
I don't know if it is and don't think that it is
your intention here, but any attempts to explain
away a human beating a bot by plain luck or by his
constantly making lesser moves just won't work...
>Hence my original theory that JF level 5 may have better
>results against a weak player than level 6 or 7.
How about rewording this as: "JF level 5 may not
be as effected from its opponents purposefully
made lesser moves than level 6 or 7 would be"...?
Obviously, at levels 5, 4, 3, 2, 1, its opponent
is expected to make lesser and lesser moves to
begin with...
>OK, the program could be refined to evaluate the top two
>or three human moves, give them weightings for the
>probability for each (i.e. the probability of a mistake),
>and proceed on that basis. Estimating the probability of
>a mistake would be very uncertain; is the computer going
>to look up your FIBS rating? Therefore I suspect that
>this is not done.
You are right that the bot can't rely on opponent's
rating. But even if it could, it still couldn't rely
on its opponents expected mistakes in any "logical"
manner. If we substitute JF's "levels" for "FIBS
ratings" and if one were able to indicate to the bot
that he's a "level 2" player but wants to play at
"level 7", the bot would then be aware of it but how
could it possibly base his moves on this info
candidly disclosed to it by the numan opponent...?
>By contrast, an expert may think, "I know remember that
>play Y is theoretically best. However, Ian has a poor
>grasp of timing considerations; let's make play X which
>tends to get us into a prime vs prime game where he'll
>make mistakes."
I agree with this as far as human vs. human games go
but this has never been a considertion for me as far
as an aspect of the issue of making purposefully
lesser moves against bots. In fact it's the opposite.
That is, you rely on the bot's always making the
best (at least statistically) moves. Let me give an
example. Early on in a game, JF rolled 11 and built
two blocks on its 7 and 5 points, with either's home
board being otherwise all open. There is nothing
unusual about this, except that I had a blot on my 5
point in a situation where I could end up blocking
him in easily also. Since I had wondered for a moment
whether I would have played it the same way without a
second thought if I were on the opposite side, I went
back to that position and experimented with it. First
I switched sides and played a few variations of games
by hitting the blot on the 5 point, which resulted in
relatively easy wins and even a gammon. Then I wanted
to see what would it take for JF to play 11 in that
situation any differently. I tried to bait him with
blots, etc. many different ways but nothing changed
its mind unless I gave him two blots to hit. Based
on this, if I were in a similar situation where the
possibility of my opponent rolling 11 would concern
me, against JF I would feel quite safe making a
variety of moves counting on the way it will play.
If I play 100,000 games against JF, its consistently
playing the 11 that way in that particular position
may eventually pay off. But, even then, how many
times that position would occur in 100,000 games...?
Let's say it would occur 100 times. Given that 11 in
that position could be played many different ways,
let's say that 5 different ways of playing it produce
the "best" results as 6, 12, 22, 28, 32 (percent of
the time). With these imaginary numbers, would a
certain move producing better results only 4 more
times than the next closest one, really significant?
What if at 50,000 games mark the numbers were (in the
same order as percentage) 8, 4, 26, 40, 22? And what
if I play 104,000 games (enough for this position to
occur 4 more times, as it would be statistically
expected according to the above imaginary ratio) and
the number 4 choice goes up to 32 to equal the "best"
move as of 100,000 games...?
Now, if I play only 1,000 games against JF and play
11 as in the number 4 way/choice, at the one and only
occurrence of that position in question above, and if
I win that game, would you tell me that I made an
inferior move and that if I play 60,000 games against
JF it will eventually beat me by 4 points (just based
on the 60 expected occurrence of this position and 11
with JF always making the "best" move and me always
making the "second-best" move"...? This is just plain
nonsense to me and I would never base a "strategy" on
some statistics which a certain robot obtained while
playing 100,000 or 1 gazillion games against itself...
What really surprizes me is that there are so many
assumedly bright people who consider a "one-track
strategy" based on dubious statistics as the only
way to go and that they find it hard to believe a
humans can win against such a strategy.
I'm not sure if I was able to express myself clearly
but I personally think that this is a very important
point and would appreciate if people interested in
this topic make a little extra effort to make sense
out of what I tried to explain.
MK
Murat Kalinyaprak
>Murat Kalinyaprak writes:
>>Here is your challenge of the day: Find and present some
>>cases where world-class players played against JF for
>>$100/point, using JF's automatic dice rolling.
>Without commenting on Murat's many other points, this is
>very easy to understand. Jellyfish's rolls are determined
>by the seed and counter. If we have Jellyfish determine
>the rolls, then the seed and counter could be set to
>produce rolls that favor one side or the other.
If dice rolls starting at seed/counter=X favors one
player, starting at seed/counter=X+1 would favor
the other side. If what you mention here could
be an issue (I'm not saying it is), then this
would mean taking a %50 chance right of the bat
by either side (which wouldn't leave much for the
300 games to be played, but at least it would still
be fair:)...
>Who gets to determine the seed and counter,
Players can take turns to roll a single dice 4
times, producing a sequnce like 5, 3, 2, 6,
which could be used to set the seed=5326. This
would have to be limited to 4 rolls since
seeds greater than 65535 cause problems:). But
since they didn't know it then, if they had hit
a certain seed/counter and JF started rolling
84's, 112's, -44's, etc. it would have been a
*world-class bug* instead of a bug announced
by an amateur in a newsgroup where some JF
groupies were too eager to trivialize... :)
>and how can this person be trusted by both sides?
The same way they trusted him to roll the dice
manually.../?
> If the seed and counter were noted, there is a tiny
> possibility that the side opposing the computer might
> have some fore-knowledge of the coming rolls.
This is humanly impossible (assuming that the
opponents were on their own and face-to-face).
>In any event, a tedious verification process would be
>necessary after a game was played to ensure that JF
>generated the rolls expected with that seed and counter.
There is no need to respond to this other than by
saying that convenience wasn't an issue, otherwise
it would misplace the focus. Please see my remarks
about this further down.
>In short, because JF's dice rolls are demonstrably completely
>deterministic, they are inappropriate for gambling. For
>a user playing against the computer for fun and education,
>with no motivation to cheat, JF's rolls are fine.
"Motivation to cheat" on whose part...? Obviously,
it can't be the human player in this situation.
And the fact that JF's dice being predetermined
is offered as one proof of Jf's not cheating by
its developer and its advocates. So, there should
be no problem with this at all. No human can
remember X million dice rolls and JF knows them
all ahead of time but claimed to not make use of
that knowledge in how it plays any given position.
The only thing that would matter is whether the dice
rolls are random. And this seems to be one question
still unanswered! by anybody based on any proof...
>Also, even if JF's dice roller is completely fair, the
>program could be modified by a dedicated cheater so
>that the program would cheat for one side or the other.
And who would be that "dedicated cheater"? Those
"world-class" players surely wouldn't attribute
such a label to JF, and JF's backers surely wouldn't
attribute such a label to those players, would they?
>With large amounts of money on the line, you want to
>guard against this possibility.
I'm glad to see you find the issue worthwhile to
contemplate further and make comments. I think the
possible problems you pointed out could be easily
overcome by using a fresh copy of JF and rolling
dice to determine the starting seed/counter. But
there was no attempt made in that direction and this
is very important because of the reason which was
given for rolling dice manually.
Remember that I had only quoted these 4 lines from
Chuck's article:
"Malcolm Davis,... invited two other world class players
"(Mike Senkiewicz and Nick Ballard) to each play 300 games
"... Dice were rolled manually (by Davis, I believe) so that the
"typical "JF cheats" arguments would be completely irrelevant.
My previous argument focused on the last statement,
which gives *the reason* (at least in Chuck's words).
Had they said they rolled manual dice because it
was more convenient, it would have been a different.
story. In fact, they could have offered any other
reason as trivial as their liking the sound of the
rattling dice. But they didn't do that, and instead
they said it was to "eliminate JF cheats arguments"!
What they have done certainly didn't do the least
bit of good to eliminate JF cheats arguments. To
the contrary, while there were other things (I would
even say more convenient things/ways) they could do
to accomplish more in that direction, they didn't do
those. Thus, I think that what logical deduction can
be derived from this is very clear in my opinion...
MK
Murat Kalinyaprak
>Murat Kalinyaprak writes:
>> Often times the difference between the "best move" and
>> the "next best move" (or even the "third/fourth/etc.
>> best move" is statistically minute. Yet, even if the
>> difference is only 1%, we know that a bot will prefer
>> it consistently. This is fine and over X million games,
>> that 1% may be visible/significant and eventually may
>> pay off. But it may not make much difference in 100 or
>> even 1000 games.
>I disagree. By "1%", I assume you mean 1% of the cubeless
>probability of winning.
Since I don't know what this really means, I can't
say whether we are talking about the same thing.
But, what I had meant was 1% difference between
"move A" and "move B", whatever the unit of measure.
For example, if they were rated on a scale of 1 to
100, one move could be rated at 100 (best move) and
another one at 99 (second-best move), etc.
>If you consistently make an error of this magnitude just
>once in every game, then you would expect to lose about 4
>points per 100 games (because the average points won per
>game is about 2), and about 40 points per 1000 games.
In case we were talking about the same thing, I
wound find this hard to believe (although may be
valid). In case we were not talking about the same
thing, then of course, it wouldn't apply. If my
clarification above helped any, maybe you can make
a follow-up comment on this.
>Now, Murat -- what would you do in this position? If you
>would be careful to select the safe play, then I claim
>that you _are_ sensitive to errors of 1%. Don't be too
>critical of others arguing over small differences, when
>you do the same yourself.
Maybe I shouldn't be critical of what others do
regardless, but there may be a misunderstanding
here about how sensitive I am. I thought I was
trying to make the point that I'm not really all
that sensitive. When I look at a few possible
moves for a given position and dice roll, I may
see "worse-bad-good-best" moves and that's about
the extent of it. Then there are basically two
possibilities and two arguments that go along
with them. One: what I think to be the best move
may be the second or third best statistically.
With this, I'm saying that if the player is good
enough, he won't choose a drastically worse move
(unless on purpose) and that consequences of very
small differences won't be visible in the short
term. Two: what I think to be the best, second
best, etc. moves may coincide with statistically
best, second best, etc. moves. With this, I'm
saying that if a player is good enough, he may
use a slightly lesser move as part of an overall
strategy and produce positive results in the
short *and* long term (if the move doesn't work,
he can still handle its consequences consistently
well enough to not get demolished by them). In
addition, I'm arguing that one may be able to
accomplish this even more easily/successfully
against a robot, since one would know (at least
to a great extent) how it will play the next
possible dice combinations.
I think the difference between probability and
backwards statistics should also be stressed.
The second one is meaningless/useless for me.
Given a certain position and a certain dice roll,
I would never take into consideration what move
produced better results over 100,000 or 1,000,000
past games. Game #100,001 may/will change the
statistics but not the probabilities. I wouldn't
mind getting your and other's views on this.
MK
Gary Wong
> >> the "next best move" (or even the "third/fourth/etc.
> >> best move" is statistically minute. Yet, even if the
> >> difference is only 1%, we know that a bot will prefer
> >> it consistently. This is fine and over X million games,
> >> that 1% may be visible/significant and eventually may
> >> pay off. But it may not make much difference in 100 or
> >> even 1000 games.
>
> >probability of winning.
>
> Since I don't know what this really means, I can't
> say whether we are talking about the same thing.
> But, what I had meant was 1% difference between
> "move A" and "move B", whatever the unit of measure.
> For example, if they were rated on a scale of 1 to
> 100, one move could be rated at 100 (best move) and
> another one at 99 (second-best move), etc.
OK, to be precise about what I mean, the glossary at Backgammon Galore
(http://www.bkgm.com/glossary.html) defines the cubeless probability
of winning as:
CPW
Cubeless Probability of Winning -- a player's chances of winning the game
if no doubling cube is being used.
Essentially, you play the rest of the game out with no cube and ignore
gammons, and estimate the probability of winning. For instance, if you
are a slight favourite, your CPW might be 55%. Your CPW obviously varies
depending what moves you make -- you can imagine being in a position where
if you make the "best" move, your CPW could be 73%, but by making a slightly
worse move you might have only 72%. This is exactly what I mean by the
"1% error" we're discussing here.
> >If you consistently make an error of this magnitude just
> >once in every game, then you would expect to lose about 4
> >points per 100 games (because the average points won per
> >game is about 2), and about 40 points per 1000 games.
>
> In case we were talking about the same thing, I
> wound find this hard to believe (although may be
> valid). In case we were not talking about the same
> thing, then of course, it wouldn't apply. If my
> clarification above helped any, maybe you can make
> a follow-up comment on this.
OK. First, I claim that the average number of points won per game is
about 2 (based on statistics from old games of mine) -- it's certainly
more than 1, because of the cube and gammons. There are a large number
of single point wins (drops of initial doubles), a moderate number of
wins of 2 points (gammons or doubled games), and smaller numbers of wins
of 3 (backgammons) and 4 points and above. Overall, if you add up all
the wins and average them, a typical game is worth about 2 points.
Now, if you and I were to play 1000 games and you won all of them,
the best estimate is the final score would be +2000 to you (1000 games
at 2 points per game on average). If we were exactly equal, then on
average I would win 1000 points (500 games x2) and you would win 1000
too, for a total score of 0. If we are otherwise equal, but I make a 1%
error in every game, then my CPW is 49% and your CPW is 51%: over 1000
games, I would expect to win 490 games (980 points) and you would expect
to win 510 games (1020 points) -- overall, you're ahead by 40 points. This
is exactly what I meant when I wrote "40 points per 1000 games" above.
> >Now, Murat -- what would you do in this position? If you
> >would be careful to select the safe play, then I claim
> >that you _are_ sensitive to errors of 1%. Don't be too
> >critical of others arguing over small differences, when
> >you do the same yourself.
>
> Maybe I shouldn't be critical of what others do
> regardless, but there may be a misunderstanding
> here about how sensitive I am. I thought I was
> trying to make the point that I'm not really all
> that sensitive.
I apologise if I misunderstood you. What I meant to imply is that I think you
_are_ sensitive to errors as small as 1%. Not because you look at two moves
and think "aha! move A wins me 76%, but move B wins only 75% -- therefore
I select move A"; humans don't play like that. What I mean is that by looking
at the two moves, you would be able to reason "move A is better than move B
because (blah blah blah)", and in many situations you will be able to do this
reliably even if the difference between the moves is 1% or smaller.
An example of the 1% error I described in a previous article is:
+13-14-15-16-17-18-------19-20-21-22-23-24-+
| O | | X X X X X X | Cube: 2
| | X | X X X X X X |
| | | X X |
| | | |
| | | |
v| |BAR| | O to play 21, money game.
| | | |
| | | |
| | | |
| | | O O O O O O |
| O O | | O O O O O O |
+12-11-10--9--8--7--------6--5--4--3--2--1-+
Let me give you two choices: 14/11, or 9/6. Which would you play? If
you chose 14/11, I claim you just made a 1% error: if you roll 66 next
turn, you'll be forced to leave a blot on the 6 point. The chance of
rolling 66 and then having the blot hit is 11/1296 (roughly 1%), and
this will very likely cost you the game (definitely, if X holds the
cube) -- in other words, a 1% error.
The only point I'm trying to make is that most intermediate or better
players would see that 9/6 is better. They DON'T think "9/6 gives me
92% wins, 8% losses, 6% gammons; 14/11 gives me 91% wins, 9% losses,
6% gammons" (that's Jellyfish, not humans). They DO immediately see
that 9/6 is a better move without any calculation of probabilities,
even though it's `only' a 1% difference. That is what I mean when I
claim a 1% error is significant.
> I think the difference between probability and
> backwards statistics should also be stressed.
> The second one is meaningless/useless for me.
> Given a certain position and a certain dice roll,
> I would never take into consideration what move
> produced better results over 100,000 or 1,000,000
> past games. Game #100,001 may/will change the
> statistics but not the probabilities. I wouldn't
> mind getting your and other's views on this.
I agree entirely (although when arguing about "probability" and "statistics"
as you use them here, bear in mind that by the strong law of large numbers,
the statistics will almost surely converge to the true probability as
you increase the number of trials without bound). Another poster (I forget
who, sorry) earlier wrote something along the lines of "the statistics are
better because the move is better, not the other way around". If the
statistics are accurate (a very big "IF"!!), then one implies the other;
but in general you owe it to yourself to select the best move. If you
claim that you don't need to know the statistics to find the best move,
then I agree with you wholeheartedly; that's exactly the point I'm trying
to make. But, if you look at the statistics (and make the very big
assumption that you have good reason to believe that the numbers you have
are an accurate reflection of the true probabilities!) and disregard
differences because they're "only" 1%, then you are making a significant
error -- as bad as moving 14/11 in the example above. 1% _IS_ significant!
MJR
[snip]
> When I look at a few possible
> moves for a given position and dice roll, I may
> see "worse-bad-good-best" moves and that's about
> the extent of it. Then there are basically two
> possibilities and two arguments that go along
> with them. One: what I think to be the best move
> may be the second or third best statistically.
> With this, I'm saying that if the player is good
> enough, he won't choose a drastically worse move
> (unless on purpose) and that consequences of very
> small differences won't be visible in the short
> term. Two: what I think to be the best, second
> best, etc. moves may coincide with statistically
> best, second best, etc. moves. With this, I'm
> saying that if a player is good enough, he may
> use a slightly lesser move as part of an overall
> strategy and produce positive results in the
> short *and* long term (if the move doesn't work,
[snip]
I can't comment on your whole post, but I think you would be surprised
at how often a human thinks two plays are close when rollouts show there
is really a BIG difference. There a book called "New Ideas in
Backgammon" by Kit Woolsey and Hal Heinrich which explores exactly this
subject. (using examples which even experts misjudge significantly) The
strength of the strong neural net backgammon programs (in my opinion) is
their ability outperform most opponents in these not so uncommon
positions.
Of course you are right to say that there are many positions where the
difference between the best 2 plays is small, and so choosing the wrong
one doen't cost you very much. Some problems which people find
interesting fall into this category..they are problems not because they
are likely to cost the average player alot of equity, but because the
stlye of play between the candidate moves is so different. I think that
because humans are so used to studing problems where the correct play is
correct by only a small margin, that they tend to overlook the problems
where the correct play is significantly superior but not obviously so.
My point is that it is easy to think when choosing between 2 plays which
look close that even if the wrong move is chosen it doesn't cost that
much, when really there is a fair chance that the difference is larger
than one might think.
Dan Frank
>>> My speculation was that one got cought in these by
making the expected (i.e. good/best) moves and that
if one could see it coming, he could prevent these
sequences/patterns by making an unexpected (i.e.
lesser/wrong) move.
>>>>
I have already remrked, that after the recommended "best move"
JF gets exactly the dice to meet the tiny risk implied in it.
Consequences: if one has to often such negative experiences
when playing the "best move", he may in the long term develop
some strange reflexes.
***************************
>>> What really surprizes me is that there are so many
assumedly bright people who consider a "one-track
strategy" based on dubious statistics as the only
way to go and that they find it hard to believe a
humans can win against such a strategy.
>>>
Here is something close to the so called self fulfilling prophecies.
Who assumed (or pretended) that they are bright people (besides themselves)?
****************************
>>>Remember that I had only quoted these 4 lines from
Chuck's article:
"Malcolm Davis,... invited two other world class players
"(Mike Senkiewicz and Nick Ballard) to each play 300 games
"... Dice were rolled manually (by Davis, I believe) so that the
"typical "JF cheats" arguments would be completely irrelevant.
My previous argument focused on the last statement,
which gives *the reason* (at least in Chuck's words).
Had they said they rolled manual dice because it
was more convenient, it would have been a different.
story. In fact, they could have offered any other
reason as trivial as their liking the sound of the
rattling dice. But they didn't do that, and instead
they said it was to "eliminate JF cheats arguments"!
What they have done certainly didn't do the least
bit of good to eliminate JF cheats arguments.
>>>>>
MK is fixed on the cheating possibility -
it's understandable and he is right to react
against such attitudes, sold as arguments, as:
"I, The Big Guru, I say that it is ridiculous to
think JF cheats, even if it gets dubiously many
lucky rolls."
The discusion became uninteresting (but is perhaps
still necessary):
JF played with definitely unbiased dice well against
Senkiewicz and Ballard - that's enough proof for it's
strength (and that's the point).
Very important: both Senkiewicz and Ballard were high class
players before the apparition of JF, t.i. the thing with
the selffulfilling prophecies from above don't apply.
**********************************
>>>> I think the difference between probability and
backwards statistics should also be stressed.
The second one is meaningless/useless for me.
Given a certain position and a certain dice roll,
I would never take into consideration what move
produced better results over 100,000 or 1,000,000
past games. Game #100,001 may/will change the
statistics but not the probabilities.
>>>>>
That's very important! The difference is even essential.
60% (f.i. after the 2nd roll) winning chances are not always 60%
(f.i. 3 rolls later), while the probability of a roll is
always the same: 1Z (1/36).
I refrained a while ago to post the following for being too
provocative, but I think now is high time:
In the last time I had the occasion to think about statics and backgammon. I
came to a basic conclusion, which is for many very provocative - I'd like
to hear some opinions:
The basical aim of statistics is to
measure few to get information about many
- f.i. "prediction" for election. At Backgammon statistics works with
similar means but exactly the opposite way:
measure many to know how to act in an unique situation.
Statistics contradicts at Backgammon itself, it's so to say perverted statistics.
If the tool is perverted, then there must be something dubious about the
results.
NB Net distinction to probabilities (of dice rolls), which are CALCULATED and
not EMPIRICAL ESTIMATED.
--
Dan Frank
editor & publisher of ESSENTIAL BACKGAMMON
Murat Kalinyaprak
> Murat Kalinyaprak writes:
>> Since I don't know what this really means, I can't
>> say whether we are talking about the same thing.
>> But, what I had meant was 1% difference between
>> "move A" and "move B", whatever the unit of measure.
>> For example, if they were rated on a scale of 1 to
>> 100, one move could be rated at 100 (best move) and
>> another one at 99 (second-best move), etc.
> ..... depending what moves you make -- you can imagine
> being in a position where if you make the "best" move,
> your CPW could be 73%, but by making a slightly worse
> move you might have only 72%. This is exactly what I
> mean by the "1% error" we're discussing here.
I guess we are not talking about the same thing but
that they may correspond in some cases (i.e. a 1%
lesser move may result in 1% lesser CPW).
> ..... What I mean is that by looking at the two moves,
> you would be able to reason "move A is better than move
> B because (blah blah blah)", and in many situations you
> will be able to do this reliably even if the difference
> between the moves is 1% or smaller.
Ok, but what I'm saying is that even if I can reliably
select a better move I may consider the difference
inconsequential and opt for a lesser move, or that even
if can't reliably select a slightly better move, it may
not make a significant difference anyway. I'll use your
example to illustrate this further below.
> +13-14-15-16-17-18-------19-20-21-22-23-24-+
> | O | | X X X X X X | Cube: 2
> | | X | X X X X X X |
> | | | X X |
> | | | |
> | | | |
> v| |BAR| | O to play 21
> | | | |
> | | | |
> | | | |
> | | | O O O O O O |
> | O O | | O O O O O O |
> +12-11-10--9--8--7--------6--5--4--3--2--1-+
> Let me give you two choices: 14/11, or 9/6. Which would
> you play? If you chose 14/11, I claim you just made a 1%
> error: if you roll 66 next turn, you'll be forced to leave
> a blot on the 6 point. The chance of rolling 66 and then
> having the blot hit is 11/1296 (roughly 1%), and this will
> very likely cost you the game (definitely, if X holds the
> cube) -- in other words, a 1% error.
The difference I were referring to as 1% difference
would be more like the one between 9/6 and 14/12 10/9
or 14/12 9/8 or 10/8 9/8 or 10/7 or 10/9 9/7, etc.
none of which would leave a blot with a 66. It's
possible that X may dance for 2, 3 more rolls and O
may end up leaving a blot later also. In that case,
one may look back and say that playing 10/8 9/8
instead of 10/9 9/7 would have been better. But
I'm personally not capable of mentally calculating
(actually probably couldn't do it even on paper
either) or estimating such odds several rolls ahead
and pick a move based on that. I think I'm good
enough to avoid moves like 14/11 in this position (if
I'm really giving myself to it), but from among the
other moves I mentioned above, I would probably just
pick any one depending on my mood. If there is a
"measurable" difference between 10/8 9/8 and 10/9 9/7
here and if some human players or bots can derive any
visible advantage from it, I would admire them. My
believing that there is no consequential difference
between those moves may not be too far off from the
reality or it may be just a self-consolation because
I'm not capable of anything more anyway...
>> I would never take into consideration what move
>> produced better results over 100,000 or 1,000,000
>> past games. Game #100,001 may/will change the
>> statistics but not the probabilities. I wouldn't
>> mind getting your and other's views on this.
> I agree entirely (although when arguing about "probability"
> and "statistics" as you use them here, bear in mind that by
> the strong law of large numbers, the statistics will almost
> surely converge to the true probability as you increase the
> number of trials without bound).
And I agree with you on this. I just don't see a
need or benefit for bots to take the "backwards"
way of going via statistics while they can go
directly via probability. It may take a human 20
years of playing to accumulate enough experience
(i.e. statistics) approaching true probability,
but a computer can be programmed to be already
there the minute it's compiled and ready to play
its very first game. Since the probabilities are
already known, what more is there for a bot to
learn from statistics/experience of playing a
million games? Human have to go that route because
we have no choice. Although our brains are much
more complex and capable in general, we just don't
have the raw number cruching power of computers.
MK
Murat Kalinyaprak
>> ........ One: what I think to be the best move
>> may be the second or third best statistically.
>> With this, I'm saying that if the player is good
>> enough, he won't choose a drastically worse move
>> (unless on purpose) and that consequences of very
>> small differences won't be visible in the short
>> term. Two: what I think to be the best, second
>> best, etc. moves may coincide with statistically
>> best, second best, etc. moves. With this, I'm
>> saying that if a player is good enough, he may
>> use a slightly lesser move as part of an overall
>> strategy and produce positive results in the
>> short *and* long term (if the move doesn't work,
>I can't comment on your whole post, but I think you would
>be surprised at how often a human thinks two plays are
>close when rollouts show there is really a BIG difference.
>There a book called "New Ideas in Backgammon" by Kit Woolsey
>and Hal Heinrich which explores exactly this subject. (using
>examples which even experts misjudge significantly) The
>strength of the strong neural net backgammon programs (in my
>opinion) is their ability outperform most opponents in these
>not so uncommon positions.
This is a good point to which I hadn't given much/any
consideration in my previous argument. The element of
"misjudging" may be more significant than I/we think.
Has anybody tried to quantify how significant this may
be, by analyzing top level games played against bots,
etc.? Does the book you mention makes any attempts in
that direction?
MK
Phill Skelton
> In the last time I had the occasion to think about statics and
> backgammon. I came to a basic conclusion, which is for many very
> provocative - I'd like to hear some opinions:
>
> The basical aim of statistics is to
>
> measure few to get information about many
>
> - f.i. "prediction" for election. At Backgammon statistics works with
> similar means but exactly the opposite way:
>
> measure many to know how to act in an unique situation.
>
> Statistics contradicts at Backgammon itself, it's so to say perverted
> statistics. If the tool is perverted, then there must be something
> dubious about the results.
Might I suggest that you don't seem to understand statistics. The sort
of stats I have to work with every day as a physicist are exactly the
sort of thing you find in backgammon - make many measurements of
something to find the probability distribution, so that you can
make predictions.
For a single event, all you can predict is the probability of a
certain outcome - the chance of rolling '3' on a six sided die
is 1/6. But you will either roll a 3 or you will not, as you can't
get 1/6th of a 3. For a large number events - 6,000,000 rolls of a die,
you would expect 1,000,000 3's, and will very likely get something
close to that if you were daft enough to do it.
Hence it is thought that you can make definite predictions about large
numbers of events, but not small numbers or single events. This isn;t
true. The same thing is really happening in both cases. But in the case
of a single roll you either get a 3 or you don't. With 6,000,000 rolls
you either get exactly 1,000,000 3's or you don't, and the chance
of getting the exact number is pretty small. But 1,000,156 is
considered close enough to 1,000,000 to fit in with the probabilities.
By the same token, with a single die 0 is close enough to 1 to
fit in with the probabilities...
So, back to backgammon. You are in the position of having to either
leave a shot 6 away from your opponent, or 5 away. If you leave
the blot 6 points away, 17/36 rolls hit it. If 5 away, 15/36 rolls
hit it. If I understand Dan correctly, he would claim that in the
unique situation in which you find yourself, your opponent may be
about to throw 4 1, or some other roll that hits on 5 but not on 6,
and in that case leaving the blot 6 away would have been the correct
move.
I would suggest that this is wrong. Unless you know in advance what
the next throw is going to be, it is always correct to leave the
fewest number of shots. I don't know if Dan's talk about human
'intuition' at backgammon as opposed to probability is a claim to
intuit that his opponent will be able to hit 5 away but not 6 away.
If no such claim is made then leaving the blot 5 away is always
the best move, even if you do get hit. The more times you leave it 6
away, the more times you will be hit, and the more games you will lose.
Phill
Claes Thornberg
to some statement about backgammon and statistics made by Dan Frank:
> Might I suggest that you don't seem to understand statistics. The sort
> of stats I have to work with every day as a physicist are exactly the
> sort of thing you find in backgammon - make many measurements of
> something to find the probability distribution, so that you can
> make predictions.
This reminds me of a story and a proposition. About 15 years ago,
Mike Svobodny visited Sweden and played in our biggest tournament.
Not only did he reach the final of that tournament, he also brought
with him a proposition which almost ruined some quite good players.
The proposition, which I'm not going to show you, was not only a take,
it was almost a beaver. But he got to play it with the usual
proposition setup, i.e. 1 point for taking the cube. He let anyone
side-bet against him, for (almost) any amount, cash after every
game. You could join games whenever you liked. E.g. you could play
one, keep out two games, play three, keep out one, etc. And now to my
point. Not only do people have trouble understanding statistics, they
have problem understanding probablity. In particular one guy, who
ultimately lost approximately $8,000 on this proposition, thought that
he could beat the odds by joing a game after a series of wins by
Svobodny. He reasoned that since Svobodny had won four, or whatever,
games in a row, the likelyhood that Svobodny would win the next game
were actually lower. But lo and behold, the punishment was if not
swift, at least sure.
Regards,
Claes Thornberg
PS. I don't know how much the guy who played the proposition
lost. He kept playing the proposition with Svobodny all through the
night. Anyway, I don't believe Svobodny left Stockholm a poorer man.
--
______________________________________________________________________
Claes Thornberg Internet: cla...@it.kth.se
Dept. of Teleinformatics URL: NO WAY!
KTH/Electrum 204 Voice: +46 8 752 1377
164 40 Kista Fax: +46 8 751 1793
Sweden
Dan Frank
Subject: Re: What is a fair test of ability?
****
**
MJR wrote:
>I can't comment on your whole post, but I think you would
>be surprised at how often a human thinks two plays are
>close when rollouts show there is really a BIG difference.
>There a book called "New Ideas in Backgammon" by Kit Woolsey
>and Hal Heinrich which explores exactly this subject. (using
>examples which even experts misjudge significantly)
**
****
FYI:
In one of the previous issues of EB I have analyzed in a
review of that book 3 of the 104 positions,
and proved that there were misplays
_according to usual backgammon-reasonment_.
Dan Frank
> Dan Frank wrote:
*****
**
>The discusion became uninteresting (but is perhaps
>still necessary):
Perhaps it became unisteresting because it became
too specific/limited to JF (which it didn't have
to be so) or perhaps calling it "uninteresting"
is a way out for some people who can't handle/keep
up with the debate... But, I also think that it's
necessary for the overall good of this newsgroup.
**
*****
Uninteresting from a practical POV: given its proven strength,
I'm interested in using it for improving skill, deepening
strategical uderstanding aso.
The dice are still a problem for the rollouts. BTW I explain the high
gammon and backgammon rate with its dice generator.
*****
**
>JF played with definitely unbiased dice well against
>Senkiewicz and Ballard - that's enough proof for it's
>strength (and that's the point).
This is "a/one point" that has been often mentioned
and constantly kept in mind during this discussion.
It has no value of proof in an argument about JF's
cheating/not cheating. Claims like "JF is already a
strong player therefore it has no reason to cheat",
etc. are only opinions/assumed conclusions and not
proof based on demonstrable facts.
**
*****
This is not a further argument in the sense of circumstantial
evidence for not cheating - interesting is now (for me -
not for you, since you have spent so much time and effort on
that matter) only the practical POV.
*****
**
>Very important: both Senkiewicz and Ballard were high class
>players before the apparition of JF, t.i. the thing with
>the selffulfilling prophecies from above don't apply.
I have no idea about whose "self-fulfilling prophecies
regarding what" that you keep referring to...?
**
*****
JF has been trained to play certain positions in a specific
way. This seem to have influence even on later training phases,
thus it estimates some specific plays higher. (There was a posting
to this problem from Brian Sheppard or David Montgomery, I think).
Now those who train with JF are used to consider those playing ways
stronger, and JF, of course, confirms this every time. If they win
in a real (against men) game, they are tempted to explain this
by the JF-estimation. So far to self-fulfilling prophecies.
In any case, at the end JF has a "personality", t.i. it prefers some
kind of plays over others, merely basing on some training particularities
(it becomes quite human <gr>!), its preferences are biased.
Senkiewicz and Ballard have their own personality, their preferences
are not biased under JF's influence (but by their own experience).
T.i. if JF could play well against them, it is indeed a strong player,
and since the dice issue has been elegantly eluded, there can be no
more doubt about his playing strength (no doubt about its primitive
interface too ...).
DF:... I'd like to hear some opinions:
MK: My humble opinions follow...
*****
**
>The basical aim of statistics is to
> measure few to get information about many
>- f.i. "prediction" for election.
This is only one usage of statistics and I doubt
that it's even a/the most common usage. Purposes
of statistics are surely not limited to this.
**
*****
Well, at least it is an important step in the determination
of data like how many hours per capita are spent to post
rubbish on the Internet in a certain coutry aso.
These data are then used to steer the economy.
*****
**
>At Backgammon statistics works with
>similar means but exactly the opposite way:
> measure many to know how to act in an unique situation.
I think this is a more common usage and is done
to come up with some probabilities where there is
no better/direct ways of doing it. For example,
it may be done to figure some randomness into
events non-random by nature or intention, like
airplane crashes. When you roll dice, you know
it will land on one of its six faces giving 1/6
probability of rolling a certain number; or when
you toss a coin, you know it will land on one
of its two faces giving 1/2 probability of head
or tail side facing up. Talking about a plane's
landing ok or crashing, we have two possibilities
but we can't say that the probability of either
happening is 1/2 (50%) because it's a controlled
event with much effort made to prevent crashing.
Nevertheless, there is a seeming randomness in
the way it happens where/when/to whom/etc. When
we say that the chances of a plane crashing is
one in a million, we are basicly constructing an
imaginary dice out of statisctics, which has a
million faces with only one of them saying "crash"
and all the rest saying "land ok". In backgammon,
there is no need for such an "imaginary dice",
since we already have "real dice"...
**
*****
For a plane huge efforts are made to prevent that 1/1 million
possible event, the crash. At backgammon one risks deliberately
30% to enhance the remaining 70% ...
*****
**
>Statistics contradicts at Backgammon itself, it's so to
>say perverted statistics. If the tool is perverted, then
>there must be something dubious about the results.
I don't see anything perverted about statistics
themselves. The question is how do you try to
make use of them. Related to "knowing how to act
in certain situation" (i.e. making a decision),
etc. they can be useful when deciding whether to
travel by plane or train for example. But, if
that decision process also involved rolling dice,
then the plane/train crash-statistics, etc. would
become useless/meaningless (or at least less so).
**
*****
I told from the beginning that it is provocative ...
Of course statistics doesn't contradicts itself, but its usage is
in a certain sense perverted. Statistics gives an average view of
something. When one uses data from the statistics in the economy he
applies them on the whole, not on an individual (though sometimes it
happens, with possible tragical consequences).
At backgammon any actual decision is an UNIQUE event.
******
**
>NB Net distinction to probabilities (of dice rolls), which
>are CALCULATED and not EMPIRICAL ESTIMATED.
I can't understand what this supposed to mean, but
I would like to add my opinion that statistics may
be a substitute for probability calculations for
humans but not needed by computers. ....
**
*****
This was not so correct, but together with the beginning:
That's very important! The difference is even essential.
60% (f.i. after the 2nd roll) winning chances are not always 60%
(f.i. 3 rolls later), while the probability of a roll is
always the same: 1Z (1/36).
it should be almost clear.
However I have to specify/correct something:
with probability I refered only to the next roll - of course probabilities
can be calculated over many rolls.
What I intended to say was that over many rolls the luck issue becomes much
more important than for a single roll (maybe not in situations "everything
or nothing"). And statistics don't pay credit to this issue ...
Dan Frank
Subject: Re: What is a fair test of ability?
.......
>I would suggest that this is wrong. Unless you know in advance what
>the next throw is going to be, it is always correct to leave the
>fewest number of shots. I don't know if Dan's talk about human
>intuition' at backgammon as opposed to probability is a claim to
>intuit that his opponent will be able to hit 5 away but not 6 away.
>If no such claim is made then leaving the blot 5 away is always
>the best move, even if you do get hit. The more times you leave it 6
>away, the more times you will be hit, and the more games you will lose.
It's obvious, that you have dearly misunderstood my posting - reread and see
the answer to MK for more.
BTW I'm impressed how good you can explain those ground things - if I'll open
a backgammon school, I'll invite you to lecture to the kindergarten-classes.
Murat Kalinyaprak
> .... Hence it is thought that you can make definite
> predictions about large numbers of events, but not
> small numbers or single events. This isn;t true.
> The same thing is really happening in both cases.
> But in the case of a single roll you either get a
> 3 or you don't. With 6,000,000 rolls you either get
> exactly 1,000,000 3's or you don't, and the chance
> of getting the exact number is pretty small. But
> 1,000,156 is considered close enough to 1,000,000
> to fit in with the probabilities.
I don't know if you are trying to overdo Dan, but
if you followed this thread from earlier articles,
the point that was raised was that you don't need
statistics when you already have means of knowing
the probabilities.
Notice that in your example you haven't rolled dice
6,000,000 times and counted 1,000,156 occurences of
3, and then concluded based on this that the chances
of rolling a 3 is 1.000156/6 (or rounded to 1/6)...
In a very very very unlikely but still possible case,
if you had rolled dice 6,000,000 times and ended up
with 2,500,000 3's, you wouldn't declare that the
the chances of rolling a 3 was 2.5/6, would you...?
When we look at a 6 faced dice, we know that if we
toss it in the air it will land on one of its 6 faces
and that the chances of the side with the number 3
facing up is 1/6. We don't need to roll dice 600,
6000 or 6000000 times in order to know that... And
if statistics didn't follow the expectations in this
case, we still wouldn't adjust probabilities based
on statistics retroactively...
If you had a very unbalanced, very irregular shaped
dice, then your only choice would be to roll it X
many times, keep the statistics and derive from them
that the chances of rolling a 3 with that particular
dice is 2.7 in 6 or whatever, but not with fairly
well balanced and proportionately faced common dice...
> I would suggest that this is wrong. Unless you know in
> advance what the next throw is going to be, it is always
> correct to leave the fewest number of shots. I don't
> know if Dan's talk about human 'intuition' at backgammon
> as opposed to probability is a claim to intuit that his
> opponent will be able to hit 5 away but not 6 away. If
> no such claim is made then leaving the blot 5 away is
> always the best move, even if you do get hit.
How did "intuition" get into this? I believe the
issue was "statistics vs. probabilities" and Dan's
comments sounded like he was following up to that
in agreement. But just to be safe and to speak for
myself, the argument is that "statistics" is not
needed to know that leaving a blot 5 away is better
than leaving a blot 6 away...
From the way you started your article, it sounds
like you know quite about statistics. If so, maybe
you can steer the discussion in a direction about
how statistics can be used in backgammon in truely
beneficial ways and try to write a little more
meaningful stuff...
MK
Phill Skelton
>
> Phill Skelton wrote:
>
>
> I don't know if you are trying to overdo Dan, but
> if you followed this thread from earlier articles,
> the point that was raised was that you don't need
> statistics when you already have means of knowing
> the probabilities.
Okay, I seem to have misunderstood Dan's point here.
Sorry.
> In a very very very unlikely but still possible case,
> if you had rolled dice 6,000,000 times and ended up
> with 2,500,000 3's, you wouldn't declare that the
> the chances of rolling a 3 was 2.5/6, would you...?
Well, that's 2,500 standard deviations away, so I guess it would be
far more likely that the die was defective in some way...
> When we look at a 6 faced dice, we know that if we
> toss it in the air it will land on one of its 6 faces
> and that the chances of the side with the number 3
> facing up is 1/6. We don't need to roll dice 600,
> 6000 or 6000000 times in order to know that... And
> if statistics didn't follow the expectations in this
> case, we still wouldn't adjust probabilities based
> on statistics retroactively...
Depends. No dice is ever going to give you *exactly* equal
chances of rolling each number. Slight imperfections will
make some numbers more likely than others by a negligible
amount. But most dice are good enough for the 1/6
assumption to be more than accurate enough.
I had a friend who nad a die that looked normal but rolled 1's
way more often than it ought to. After a while it is probably
more accurate to conclude that the die does have a biased
probability distribution, as opposed to assuming that the chance
of rolling 1 really was 1/6 and we were just phenominally unlucky.
<snip>
> I believe the
> issue was "statistics vs. probabilities" and Dan's
> comments sounded like he was following up to that
> in agreement. But just to be safe and to speak for
> myself, the argument is that "statistics" is not
> needed to know that leaving a blot 5 away is better
> than leaving a blot 6 away...
I don't see how stats vs probabilities can be an issue. When
chosing the best move, you are always dealing with probabilities.
In some cases the probabilities can be calculated exactly. In
complex positions the probabilities cannot be calculated, and
statistical methods are used to empirically estimate the
probability distribution. In the case of leaving a blot 5 away
or 6 away, it is easy to calculate the exact probability of being
hit, as you say.
someona (Murat I think) wrote:
>>> I think the difference between probability and
>>> backwards statistics should also be stressed.
>>> The second one is meaningless/useless for me.
>>> Given a certain position and a certain dice roll,
>>> I would never take into consideration what move
>>> produced better results over 100,000 or 1,000,000
>>> past games. Game #100,001 may/will change the
>>> statistics but not the probabilities.
Fine as long as the probabilities are known exactly.
100,000 games will tell you quite a bit about the probability if
you have no other way of knowing it. You never find the exact
numbers, but as more games are considered, the statistics converge
slowly towards the theoretical probabilities. After all, the
difinition of probability is as the limiting case of an infinite
number of trials - at which point the uncertainty in the statistics
becomes zero. Shame you never get there.
The probability is the fundamental concept, and defines what the
best move is. I *think* that this is what you are saying as well.
The statistics reflect the probabilities to a certain degree
of accuracy (depending on the number of trials). If you can
estimate or calculate the probabilities to a better accuracy
than the available statistics give you, then go with the more accurate
option.
Phill
Murat Kalinyaprak
>Murat Kalinyaprak wrote:
>> Dan Frank wrote:
>>JF played with definitely unbiased dice well against
>>Senkiewicz and Ballard - that's enough proof for it's
>>strength (and that's the point).
>This is "a/one point" that has been often mentioned
>and constantly kept in mind during this discussion.
>It has no value of proof in an argument about JF's
>cheating/not cheating. Claims like "JF is already a
>strong player therefore it has no reason to cheat",
>etc. are only opinions/assumed conclusions and not
>proof based on demonstrable facts.
>This is not a further argument in the sense of circumstantial
>evidence for not cheating - interesting is now (for me -
>not for you, since you have spent so much time and effort on
>that matter) only the practical POV.
I'm not dwelling on this because whether JF cheats
or not matters all that much to me, but because I
believe whoever requires valid/logical arguments
form others has to provide the same. I hope this
is clear to most readers by now.
Your comment that "JF has done well aginst Senkiewicz
and Ballard" could be reworded by some as "Senkiewicz
and Ballard have done well against JF". In that case,
the question that arises is whether they would have
done just as well against JF with JF rolling the dice.
If you are going to answer this by saying "Of course,
JF's dice is no different than rolling manual dice",
then you could perhaps take your turn to relate from
them or speculate on your own as to why they haven't
played against JF with JF rolling the dice...?
>I have no idea about whose "self-fulfilling prophecies
>regarding what" that you keep referring to...?
>JF has been trained to play certain positions in a specific
>way. This seem to have influence even on later training phases,
>thus it estimates some specific plays higher....
Thanks for clarifying. I agree on this.
>For a plane huge efforts are made to prevent that 1/1 million
>possible event, the crash. At backgammon one risks deliberately
>30% to enhance the remaining 70% ...
My point on this issue is that we calculate that
"risk" in backgammon based on probabilities, not
statistics. It looks like we agree on that.(?)
>What I intended to say was that over many rolls the luck issue
>becomes much more important than for a single roll (maybe not
>in situations "everything or nothing"). And statistics don't
>pay credit to this issue ...
For as much as I understand what you mean, let me
approach this from the "flaw in the dice rolls"
angle. At 1/36 odds, one would expect a certain
specific dice to occur maybe about 2 times during
a game of average legth (60-80 rolls?). In 100
games, one would expect about 200 double 6's for
example. If we get 192, 208 or even 250 we may not
make too much of it. But if we get 400 of them, we
may suspect something wrong with the physical dice
(or the dice roller algorithm of a bot.
The odds of rolling 3 specific combinations in a
row is exponentially smaller. For example, 22+66+53
would be expected [[1/36]/36]/21 of the time (one
in about 27,000?). At let's say 70 rolls per game,
one would expect to see this about once every 400
games. In 100 games, if we see it twice, that would
be 2x more significant/noticeable than seeing 400
double 6's in 100 games. Even this happening twice
in 100 games may not be enough to make too much out
of it, but seeing it happen 3 times would probably
be enough to make one again wonder whether there's
something wrong with the dice or the roller...
MK
Murat Kalinyaprak
>Murat Kalinyaprak wrote:
>> I don't know if you are trying to overdo Dan, but
>> if you followed this thread from earlier articles,
>> the point that was raised was that you don't need
>> statistics when you already have means of knowing
>> the probabilities.
>Okay, I seem to have misunderstood Dan's point here.
>Sorry.
No problem, as long as we all understand each other
at the end and some good comes out of it, it may be
worth our efforts.
>> In a very very very unlikely but still possible case,
>> if you had rolled dice 6,000,000 times and ended up
>> with 2,500,000 3's, you wouldn't declare that the
>> the chances of rolling a 3 was 2.5/6, would you...?
>Well, that's 2,500 standard deviations away, so I guess
>it would be far more likely that the die was defective
>in some way...
With the disclaimer that we can't conclude for sure,
I would conclude the same to be most "likely". But
I hope we all agree that none of us would be tempted
to "adjust" the probabilities based on such outcomes.
>I don't see how stats vs probabilities can be an issue.
>When chosing the best move, you are always dealing with
>probabilities. In some cases the probabilities can be
>calculated exactly. In complex positions the probabilities
>cannot be calculated, and statistical methods are used to
>empirically estimate the probability distribution. In the
>case of leaving a blot 5 away or 6 away, it is easy to
>calculate the exact probability of being hit, as you say.
Ok, so we are agreeing on this. And with this in mind
and regarding making decisions in backgammon, I was
thinking that they could be made based on probabilities
alone (especially by robots) and was wondering how
statistics helped (or if they were even needed at all)
in making decisions in backgammon.
>someona (Murat I think) wrote:
>>>> Given a certain position and a certain dice roll,
>>>> I would never take into consideration what move
>>>> produced better results over 100,000 or 1,000,000
>>>> past games. Game #100,001 may/will change the
>>>> statistics but not the probabilities.
>The probability is the fundamental concept, and defines
>what the best move is. I *think* that this is what you
>are saying as well. The statistics reflect the probabilities
>to a certain degree of accuracy (depending on the number
>of trials). If you can estimate or calculate the probabilities
>to a better accuracy than the available statistics give you,
>then go with the more accurate option.
When I saw the section you quoted from a previous
article of mine, I thought I was going to have
some explaining myself to do, because in an other
article, the same day or the day after, I had said
humans had no choice but to substitute statistics
for probabilities (since we are not very good at
raw number cruching as robots). But after reading
your response I was relieved that at least I didn't
have to explain myself to you (because of the seeming
contradiction). So, as humans, we compromise for a
mixture of the two. But what about robots? If you or
others care to elaborate on this, what I would be
interested in knowing/learning is how do statistics
help robots in making decisions (or for any other
purpose) in backgammon...? I feel that if I had the
mental capability to calculate the probabilities a
few moves deep, I could just be explained the rules
and object of the game and ten minutes later beat
the world-champ. Why can't this be true for a robot
that has all the number crunching power it needs...?
MK
Dan Frank
Subject: Re: What is a fair test of ability?
*****
**
Your comment that "JF has done well aginst Senkiewicz
and Ballard" could be reworded by some as "Senkiewicz
and Ballard have done well against JF". In that case,
the question that arises is whether they would have
done just as well against JF with JF rolling the dice.
If you are going to answer this by saying "Of course,
JF's dice is no different than rolling manual dice",
then you could perhaps take your turn to relate from
them or speculate on your own as to why they haven't
played against JF with JF rolling the dice...?
**
*****
I don't say that JFs dice are as good as manual, on the contrary.
But this doesn't deny JFs playing strength.
On the other hand, I neither would like to play significant games
with dice, I don't roll, the less if it's a machine which throws them.
*****
**
My point on this issue is that we calculate that
"risk" in backgammon based on probabilities, not
statistics. It looks like we agree on that.(?)
**
*****
On a short perspective, 1-2 rolls,
the players decide on calculable probabilities, like number
of hits ASO, though strategical considerations or experience or
reference positions with known values may make one to chose
not the move with the best immediat probabilities.
In cube-situations only seldomly one can decide on probabilities (when
the game/match issue depends on 1-2 rolls).
This means in fact, you can not exclude "statistics".
*****
**
>What I intended to say was that over many rolls the luck issue
>becomes much more important than for a single roll (maybe not
>in situations "everything or nothing"). And statistics don't
>pay credit to this issue ...
For as much as I understand what you mean, let me
approach this from the "flaw in the dice rolls"
angle. At 1/36 odds, one would expect a certain
....
**
*****
Nope. I said quite plainly "luck issue", t.i. at each roll one may
realize his goal _because_ as well as _despite_ the probabilities.
The longer the game lasts (more rolls) the more this "esoteric" aspect
is involved. Statistics don't consider it, in fact, a statistics (average
outcome) is best, when it doesn't include such factors - the ideal statistics
is like the perfectly calculable probability. But reality is different
(the Big Numbers Law confirms and infirms the probability: "It _exists_ a
number...").
That means, even assuming perfect statistics, it's value is the more
diminished by this "esoteric" aspect the longer the game lasts. That's why
differences between moves like 70.25% and 69.12% are not so important,
as the (as I call them) figures-fetishists pretend (t.i. decisive).
With equities - f.i. 0.367 vs. 0.353 ppg - even more figures (different
average outcomes) are involved, t.i. the more they should be seen relatively
and not absolutely.
BTW "absolutisezers" (?!...) are fanatics, don't try to be reasonable, don't
ask for rational arguments. They don't go beyond that technical aspect,
to which they stick, I doubt that they even can.
(errare humanum est, perseverare diabolicum)
(PS it seems that I agree with Phill Skelton, or viceversa, when he says
something of the kind: statistics is nothing but noncalculable probability -
of course in circumstances like backgammon.
However, I definitely don't agree, when he says "prediction": on the basis
of statistics you can make only assumptions, (pre)suppositions. One can make
predictions, only if he is absolutely sure - f.i. fortune tellers <gr>.
I guess that figures-fetishists saying "predictions" mean "predictions".)
*****
**
The odds of rolling 3 specific combinations in a
row is exponentially smaller. For example, 22+66+53
would be expected [[1/36]/36]/21 of the time (one
in about 27,000?). At let's say 70 rolls per game,
one would expect to see this about once every 400
games. In 100 games, if we see it twice, that would
be 2x more significant/noticeable than seeing 400
double 6's in 100 games. Even this happening twice
in 100 games may not be enough to make too much out
of it, but seeing it happen 3 times would probably
be enough to make one again wonder whether there's
something wrong with the dice or the roller...
**
*****
That's why I have inventend the Baffle Cup (tm).
Murat Kalinyaprak
>Murat Kalinyaprak wrote:
>> Your comment that "JF has done well aginst Senkiewicz
>> and Ballard" could be reworded by some as "Senkiewicz
>> and Ballard have done well against JF". In that case,
>> the question that arises is whether they would have
>> done just as well against JF with JF rolling the dice.
>> If you are going to answer this by saying "Of course,
>> JF's dice is no different than rolling manual dice",
>> then you could perhaps take your turn to relate from
>> them or speculate on your own as to why they haven't
>> played against JF with JF rolling the dice...?
> I don't say that JFs dice are as good as manual, on the contrary.
> But this doesn't deny JFs playing strength.
> On the other hand, I neither would like to play significant games
> with dice, I don't roll, the less if it's a machine which throws them.
*WHY*...?! This is the question that the issue stands
on.
We already know for fact that they didn't play against
JF with JF's rolls but we don't know *WHY*...
Now we also know for fact that you won't play against
a machine unless you roll the dice manually (i.e. if
you were playing against JF which is a "machine", you
wouldn't play with its rolls) but we don't know *WHY*
(at least not exactly/explicitly)...
I can understand that you wouldn't want to speculate
on the *WHY* part for other people, but can you at
least tell us *WHY* "you" wouldn't (i.e. against JF)...?
After you answer that, please tell me if the fact that
JF is a strong player would make a difference in your
decision (i.e. would it be of any value to change your
mind about not letting JF roll the dice)...? If the
answer is "no", then please tell me *why* do you keep
referring to JF's playing strength as though it was
relevant to the issue here...?
And one last question before I finish: *why* is there
so much tail-chasing in this newgroup...? Can we move
on and make some progress please...
MK