Optimizing Gnu BG speed
tanglebear
other bots set at their highest setting. But maybe I don't know how to
set it up correctly. What I did is go into Settings|Players and set ply
for move and cube to 3 ply. I've also tried different levels of Reduced
Evaluations. I have not tried changing Move Filter or Noise. Any
advice?
Scott Steiner
Setting up 3 ply for both checker play and cube decisions will put your
patience at a test. I personally play gnubg at the preset expert level
for both checker play and cube decisions - it plays a very good game and
the response is instant. I analyse my matches at 2ply for checker play
and 3ply for cube decisions.
Nis Jorgensen
wrote:
You should note that gnubg 3-ply equals Snowie 4-ply - for some reason
they are counting in different ways. Also, as you probably noticed,
gnubg does not have any "highest settings" - you can set it to do
64-ply if you like.
--
Nis Jorgensen
Amsterdam
Please include only relevant quotes, and reply below the quoted text. Thanks
jthyssen
> You should note that gnubg 3-ply equals Snowie 4-ply - for some reason
> they are counting in different ways.
An explantion:
The reason behind this is that we name a static neural net evaluation
0-ply, because there is no look-ahead. Similarly, we name the 2-ply
evaluation "2-ply" because gnubg looks two full rolls ahead. I can't
see the logic behind naming a two roll lookahead evaluation for
"3-ply" :-)
You can argue that you count "moves" instead of "rolls", so in a 2-ply
chequer evaluation you look three moves ahead, hence it makes sense to
call it 3-ply. However, in a 2-ply cube evaluation you only look two
moves ahead, so it really doesn't make any sense to call it "3-ply".
Jørn
Douglas Zare
>Nis Jorgensen <n...@dkik.dk> wrote in message news:<ib0u7vco0c6u8bjep...@4ax.com>...
>
>>You should note that gnubg 3-ply equals Snowie 4-ply - for some reason
>>they are counting in different ways.
>>
>>
>An explantion:
>
>The reason behind this is that we name a static neural net evaluation
>0-ply, because there is no look-ahead. Similarly, we name the 2-ply
>evaluation "2-ply" because gnubg looks two full rolls ahead. I can't
>see the logic behind naming a two roll lookahead evaluation for
>"3-ply" :-)
>
>You can argue that you count "moves" instead of "rolls", so in a 2-ply
>chequer evaluation you look three moves ahead, hence it makes sense to
>call it 3-ply.
>
Ah, so you do see the logic, whether or not you agree with it.
>However, in a 2-ply cube evaluation you only look two
>moves ahead, so it really doesn't make any sense to call it "3-ply".
>
>
Isn't the real explanation that you don't care about standardizing
terminology, rather than that you also have a better definition of
cubeless equity? For the evaluations of moves, TD-Gammon, Jellyfish, and
Snowie used the same definitions of these terms long before gnu. That
should count for something, no?
I suggest that gnu change. It hinders communication if you make up new
definitions of established terms, introducing ambiguity, and I don't see
any real justification for this.
Douglas Zare
Tom Keith
While I'm at it, sign me up as favoring zero-origin array indexing
and little-Endian byte ordering.
Tom
Douglas Zare
Tom Keith wrote:
> I prefer Gnu's terminology. It makes more sense.
>
> While I'm at it, sign me up as favoring zero-origin array indexing
> and little-Endian byte ordering.
Well, I think the word "unique" should mean that there is only one. However, it doesn't, according to
the dictionary and popular usage. As a mathematician, I occasionally get requests from people who want
to know official terms for things like a 14-gon. I reply that if you say 14-gon in geometry, everyone
knows what you mean, but if you say tetrakaidecagon, only a handful of people will understand you. My
9th grade English teacher mentioned that the word "gonnorrhea" is one of the English words that sounds
the most beautiful, except for the meaning, of course. If you decide that such a beautiful word should
be reclaimed for more pleasant purposes, you simply won't be communicating.
As for plies, what people call higher-ply evaluations are really truncated rollouts. A real n-ply search
would minimax over all of the legal moves, not just those recommended by 1-ply evaluations. I don't
recommend correcting the common usage to exclude the current main use. I don't recommend changing the
definition of 1-ply or of cubeless equity just to show that they could be defined another way.
Douglas Zare
jthyssen
> jthyssen wrote:
>
> >The reason behind this is that we name a static neural net evaluation
> >0-ply, because there is no look-ahead. Similarly, we name the 2-ply
> >evaluation "2-ply" because gnubg looks two full rolls ahead. I can't
> >see the logic behind naming a two roll lookahead evaluation for
> >"3-ply" :-)
> >
> >You can argue that you count "moves" instead of "rolls", so in a 2-ply
> >chequer evaluation you look three moves ahead, hence it makes sense to
> >call it 3-ply.
> >
> Ah, so you do see the logic, whether or not you agree with it.
Please regard this and the previous posting as my personal opinions.
They are not necessarily the opinions of the rest of the gnubg
developers.
I wouldn't use the word "logic" to describe a scheme which is far from
logical.
I should add that the numbering of plies in gnubg wasn't invented by
me.
> >However, in a 2-ply cube evaluation you only look two
> >moves ahead, so it really doesn't make any sense to call it "3-ply".
> >
> >
> Isn't the real explanation that you don't care about standardizing
> terminology, rather than that you also have a better definition of
> cubeless equity?
> For the evaluations of moves, TD-Gammon, Jellyfish, and
> Snowie used the same definitions of these terms long before gnu. That
> should count for something, no?
Yes, but I don't want to support what I think is a bad standard: you
count moves for chequer play decisions and moves+1 for cube decisions.
I really can't see the logic behind this.
For example, somebody might consider MS Word documents as the standard
for exchanging documents. Again, I do not want to support what I
personally believe is a bad standard!
>
> I suggest that gnu change. It hinders communication if you make up new
> definitions of established terms, introducing ambiguity, and I don't see
> any real justification for this.
It would be quite interesting to hear or read why they originally
chose this numbering scheme! Do you happen to have any references to
the first mentioning of the term "plies" or "ply"?
Personally I do not think this should be changed in gnubg, but of
course, if the majority of the developers think it should, then they
can do it! I won't (or can't) hinder it, but I won't help either!
Joern
Tom Keith
ply evaluation" seems to say one thing but actually means another.
A static evaluation of a position involves no lookahead, so it
should be called a zero-ply evaluation.
As far as move selection goes, you could argue that the process
of choosing among several plays is doing some lookahead, but is
it a full ply of lookahead? One turn in (cubeless) backgammon has
two parts: (1) the roll of the dice and (2) the play of the checkers.
When selecting a move, you have already rolled the dice so you are
really only playing part of your turn; you're not really doing a
full ply of lookahead (which would involve looking at each possible
play for each of the 21 possible rolls).
I guess the problem is that we are trying to take definitions of terms
from other games and apply them to backgammon where they don't quite fit.
Tom
Douglas Zare
jthyssen wrote:
> Douglas Zare <za...@math.columbia.edu> wrote in message news:<3E80355E...@math.columbia.edu>...
> It would be quite interesting to hear or read why they originally
> chose this numbering scheme! Do you happen to have any references to
> the first mentioning of the term "plies" or "ply"?
Within backgammon, I have references in front of me going back to 1991 (_Learning from the Machine:
Bill Robertie versus TD-Gammon_ copyright 1993 but about a session from 1991). I presume that a bit
more work would turn up slightly earlier backgammon references, but of course in other games it has
been used much longer.
> Personally I do not think this should be changed in gnubg, but of
> course, if the majority of the developers think it should, then they
> can do it! I won't (or can't) hinder it, but I won't help either!
I think it simply adds confusion. Similarly, ignoring gammons in calculating cubeless equity just
because of the Jacoby rule is absurd, and contrary to what everyone else does. Were these innovations,
or a matter of not looking up what others mean by the terms, and why?
Douglas Zare
jthyssen
> jthyssen wrote:
>
> > Personally I do not think this should be changed in gnubg, but of
> > course, if the majority of the developers think it should, then they
> > can do it! I won't (or can't) hinder it, but I won't help either!
>
> I think it simply adds confusion. Similarly, ignoring gammons in calculating cubeless equity just
> because of the Jacoby rule is absurd, and contrary to what everyone else does. Were these innovations,
> or a matter of not looking up what others mean by the terms, and why?
Perhaps there is confusion over "cubeless equity" versus "money
equity"?!
The point is that gnubg *always* plays according to score, which, I
admit, is really a bad idea for cubeless money rollouts. Cubeless and
cubeful equities are always calculated according to score.
In gnubg we use the following as the cubeless equity (ignoring
backgammons)
eq = 2 * p(win) - 1 + gp(wg) * p(wg) - gp(lg) * p(lg)
where p(win) is the game winning chance, and p(wg) and p(lg) are the
probability of winning or losing a gammon, respectively. gp(wg) and
gp(lg) are related to the gammon prices.
The problem arises in money game played with Jacoby rule and centered
cube, for which the gp(wg)=gp(lg)=0. I'm currently working on an
option to ignore Jacoby rule for cubeless rollouts. I guess when most
people do a cubeless money rollout they want to ignore the Jacoby rule
and use gp(wg)=gp(pg)=1.
For match play rollout I think you'll generally want chequerplay
according to score even though the rollout is cubeless.
I hope I've demonstrated the definition of cubeless equity used in
gnubg is an attempt to be consistent, rather than the desire of being
different or introducing confusion.
Jørn
PS! Even though gnubg does not follow tradition on all backgammon
terms, we do still use the term "rollout" instead of "Monte Carlo
Simulation" :-)
tanglebear
> You should note that gnubg 3-ply equals Snowie 4-ply - for some reason
> they are counting in different ways.
That explains a lot! I was wondering why gnubg was so slow. I just set
it down to 2-ply and the speed is okay now. Now I can use it while
figuring out the speed reduction and noise settings mean.
Walter Trice
way that it applies consistently to both move-selection and evaluation. I.
e., if you compute an evaluation directly from the position without taking
the rolls and plays into account that's 1-ply (static) evaluation.
Evaluation which takes an average over all the rolls, assuming something
(normally "best") about what play is made for each roll, is 2-ply
evaluation. Etc.
Then the process of selecting a move for a given roll based on n-ply
evaluation can be called n-ply move selection.
The sequence of things goes position->roll->play->position->roll->play and
so on. Certainly you can break this down in various different ways.
For instance you could just count the arrows. Then your static evaluation
would be 0-ply and an evaluation that took into account possible positions
after next roll and play would be 3-ply. Or you could count both the words
and the arrows, and then static (0-ply) would be 1-ply and the (formerly
3-ply) lookahead would be 7-ply.
Standard usage seems to count occurrences of "position." This works for me.
-- Walter Trice
"Tom Keith" <tom...@ETEbkgm.com> wrote in message
news:3E80A31F...@ETEbkgm.com...
Tom Keith
>
> I don't see that there's a problem, since you can define "n-ply" in such a
> way that it applies consistently to both move-selection and evaluation. I.
> e., if you compute an evaluation directly from the position without taking
> the rolls and plays into account that's 1-ply (static) evaluation.
> Evaluation which takes an average over all the rolls, assuming something
> (normally "best") about what play is made for each roll, is 2-ply
> evaluation. Etc.
Ok, I can see that both definitions of "n-ply evaluation" work.
Would you agree then with the following usage:
Snowie's 1-ply (static) evaluation looks ahead zero plies.
Snowie's 2-ply evaluation looks ahead one ply.
Snowie's 3-ply evaluation looks ahead two plies.
Gnu's 0-ply (static) evaluation looks ahead zero plies.
Gnu's 1-ply evaluation looks ahead one ply.
Gnu's 2-ply evaluation looks ahead two plies.
If everyone agrees on this use of the word "ply" to measure the degree
of lookahead, we at least have an unambiguous (if slightly awkward) way
of describing the depth of evaluation in a backgammon program.
I'm reminded of the difference in how Europeans and North Americans count
floors in buildings. In North America if you start on the Ground Floor
and walk up one flight of stairs you end up on the Second Floor. In Europe
if you start on the Ground Floor and walk up one flight of stairs you end
up on the First Floor. But climbing one flight of stairs makes you equally
tired no matter what side of the ocean you do the climbing.
Tom
David Montgomery
<za...@math.columbia.edu> wrote:
>Isn't the real explanation that you don't care about standardizing
>terminology, rather than that you also have a better definition of
>cubeless equity? For the evaluations of moves, TD-Gammon, Jellyfish, and
>Snowie used the same definitions of these terms long before gnu. That
>should count for something, no?
I do not believe there has ever been anything like the sort of
consensus you are portraying here. I know that when Snowie came
out I felt that they had added 1 to all the ply for marketing reasons.
I have handy my JellyFish 2.0 manual. Let me quote what it says
for level 7:
"Level 5 here makes a longer list of candidates, and these are
evaluated with a combination of lookahead and simplified double
lookahead (2-ply). This means that some of the 2-roll sequences
are analyzed. Doing this calculation on a list of positions can
be considered an extra level of looking ahead, and could therefore
be called '3-ply play'. This level plays close to perfect
backgammon, and is meant for analysis of critical positions.
It is playable on a Pentium."
This certainly isn't so clear-cut in favor of Snowie's terms.
I find stronger support here for calling level 7 2-ply ("double
lookahead", "2-ply"). The 3-ply definition is hedged with "can be
considered", "could be called", and the quoting of '3-ply play'.
And in any event it would only apply to checker play.
>I suggest that gnu change. It hinders communication if you make up new
>definitions of established terms, introducing ambiguity, and I don't see
>any real justification for this.
My perception was and is that Snowie made up a new definition,
introducing ambiguity, without any justification except marketing.
For what little it is worth, I agree with the gnubg terminology,
and I suggest that Snowie change. :-)
If I were preparing to market a backgammon product, however,
(as you are), I would almost certainly use Snowie's definitions.
(After writing this, I searched a bit in the rec.games.backgammon
archives, and discovered that most people used 1-ply like Snowie
does, even before Snowie came out. I feel I must concede that
this was the more widespread usage. I still prefer gnubg's,
and still believe that there was no clear-cut standard.)
David Montgomery
Paul Tanenbaum
> ...
> But climbing one flight of stairs makes you equally
> tired no matter what side of the ocean you do the climbing.
Only if the latitudes and altitudes are equal.
---
Paul T, Professor of Geophysics, R.g.b. U.
Frank Berger
> My objection to the Jellyfish/Snowie definition of ply is that "one
> ply evaluation" seems to say one thing but actually means another.
> A static evaluation of a position involves no lookahead, so it
> should be called a zero-ply evaluation.
>
> As far as move selection goes, you could argue that the process
> of choosing among several plays is doing some lookahead, but is
> it a full ply of lookahead? One turn in (cubeless) backgammon has
> two parts: (1) the roll of the dice and (2) the play of the checkers.
> When selecting a move, you have already rolled the dice so you are
> really only playing part of your turn; you're not really doing a
> full ply of lookahead (which would involve looking at each possible
> play for each of the 21 possible rolls).
>
> I guess the problem is that we are trying to take definitions of terms
> from other games and apply them to backgammon where they don't quite fit.
If I have to move it's a decision so it's a ply. The other moves, so
it's another decision and the 2nd ply. I regard the dice as external
influence, irrelvant to ply counting.
IMHO this fits perfectly with the usage in chess and checkers and
seems to be natural to me.
> Douglas Zare wrote:
> > As for plies, what people call higher-ply evaluations are really truncated
> > rollouts. A real n-ply search would minimax over all of the legal moves,
> > not just those recommended by 1-ply evaluations. I don't recommend
> > correcting the common usage to exclude the current main use. I don't
> > recommend changing the definition of 1-ply or of cubeless equity just to
> > show that they could be defined another way.
I disagree on this definition. In chess there are also moves
truncated. The search depth is quite different for different moves and
it's called ply anayway. In BG it's just much more easy to truncated
moves....
ciao
Frank
Albert Silver
> On Tue, 25 Mar 2003 10:54:18 GMT, Douglas Zare
> <za...@math.columbia.edu> wrote:
> >Isn't the real explanation that you don't care about standardizing
> >terminology, rather than that you also have a better definition of
> >cubeless equity? For the evaluations of moves, TD-Gammon, Jellyfish, and
> >Snowie used the same definitions of these terms long before gnu. That
> >should count for something, no?
>
> I do not believe there has ever been anything like the sort of
> consensus you are portraying here. I know that when Snowie came
> out I felt that they had added 1 to all the ply for marketing reasons.
Yes, for what it's worth I find GNU's ply definition more logical. As
a chess player who has used dozens of chess engines, the term ply has
always meant how many moves it looked ahead. In fact I always thought
Snowie was looking 3 plies AHEAD until I discovered GNU. At first I
didn't even realize it was Snowie the problem, but when I did I felt
tricked, though to be fair the time difference was only in the order
of a month since I took this game up late last year. At first, I had
thought that GNU was too slow/weak to reach 3 plies (checker plays to
be clear) like Snowie. :-)))
Due to their long-term presence and widespread use, Snowie's may be
the current "standard" but it's a stupid one IMO. Reminds me slightly
of the confusion of when the new Millenium started caused by the fact
that the counting was started with one and not a zero.
Albert Silver
Douglas Zare
David Montgomery wrote:
> On Tue, 25 Mar 2003 10:54:18 GMT, Douglas Zare
> <za...@math.columbia.edu> wrote:
> >Isn't the real explanation that you don't care about standardizing
> >terminology, rather than that you also have a better definition of
> >cubeless equity? For the evaluations of moves, TD-Gammon, Jellyfish, and
> >Snowie used the same definitions of these terms long before gnu. That
> >should count for something, no?
> [...]
> "Level 5 here makes a longer list of candidates, and these are
> evaluated with a combination of lookahead and simplified double
> lookahead (2-ply). This means that some of the 2-roll sequences
> are analyzed. Doing this calculation on a list of positions can
> be considered an extra level of looking ahead, and could therefore
> be called '3-ply play'. This level plays close to perfect
> backgammon, and is meant for analysis of critical positions.
> It is playable on a Pentium."
>
> This certainly isn't so clear-cut in favor of Snowie's terms.
You are right, but I believe Jellyfish's apparent preference is the weakest
of the three. Snowie's usage clearly follows Snowie's definitions. Since
Snowie has been out for 5 years, that matters.
If gnu follows Snowie's definitions, then there is no ambiguity. If gnu does
not, then there is ambiguity every time someone says "2-ply."
> I find stronger support here for calling level 7 2-ply ("double
> lookahead", "2-ply"). The 3-ply definition is hedged with "can be
> considered", "could be called", and the quoting of '3-ply play'.
> And in any event it would only apply to checker play.
>
> >I suggest that gnu change. It hinders communication if you make up new
> >definitions of established terms, introducing ambiguity, and I don't see
> >any real justification for this.
>
> My perception was and is that Snowie made up a new definition,
> introducing ambiguity, without any justification except marketing.
> For what little it is worth, I agree with the gnubg terminology,
> and I suggest that Snowie change. :-)
I presume you don't mean for cubeless equity, too. That is a clear case of
gnu attempting to be consistent with itself instead of what others use, since
gnu's new definition is completely useless in money play.
If everyone for some reason decides to follow gnu's alternative definition,
or to call red blue, then there will be no ambiguity. That's not a reason to
change to any particular convention over any other.
> If I were preparing to market a backgammon product, however,
> (as you are), I would almost certainly use Snowie's definitions.
Indeed, internally some of Zbot's logic calls the static evaluation 0-ply,
but what the users see will not. If I say that Zbot is fast enough to play
against on 3-ply, that will just cause confusion, but if I say 4-ply more
people will understand what I mean. Still more people would understand
exactly what is meant if gnu didn't use a different definition.
> (After writing this, I searched a bit in the rec.games.backgammon
> archives, and discovered that most people used 1-ply like Snowie
> does, even before Snowie came out. I feel I must concede that
> this was the more widespread usage. I still prefer gnubg's,
> and still believe that there was no clear-cut standard.)
There would be a clear-cut standard now if gnu were to follow the general
usage, rather than trying to use a different definition.
Douglas Zare
Jørn Thyssen
[snip]
> I presume you don't mean for cubeless equity, too. That is a clear case of
> gnu attempting to be consistent with itself instead of what others use, since
> gnu's new definition is completely useless in money play.
How do you define cubeless equity for money play and match play,
respectively?
Jørn
David Montgomery
<za...@math.columbia.edu> wrote:
>David Montgomery wrote:
>You are right, but I believe Jellyfish's apparent preference is the weakest
>of the three. Snowie's usage clearly follows Snowie's definitions. Since
>Snowie has been out for 5 years, that matters.
I agree with this, mainly because JellyFish labeled things with
levels, rather than ply, for the most part (e.g., in the best
moves dialog).
After I wrote my earlier post I read more of the JellyFish manual
and remembered more clearly why I had felt that Snowie's definition
was a change from the then-standard. To get a level 7 evaluation
in JellyFish, you would select a menu item called "2 Ply". That
is pretty unambiguous.
During the mid to late 90's I must have typed control-2 (the shortcut
for that menu item) a gazillion times. I'm certain that I *never*
thought typing control-2 produced a 3-ply evaluation.
The dialog box produced by control-2 or selecting the "2 Ply" menu
item was labeled "Evaluation Level 7". Also during this time
period I must have typed control-B (for Best Moves) four gazillion
times, and then clicked "Verify all" twice. The evaluations would
then be labeled with a "7". This is not as clear as the cube
evaluation, but I know that I intuitively related this "7" to the
"Evaluation Level 7", which I knew was "2 ply". Perhaps some
people counted to 3 here, but I really think most either thought
of this as "2 ply", or didn't think about it at all.
>Indeed, internally some of Zbot's logic calls the static evaluation 0-ply,
>but what the users see will not. If I say that Zbot is fast enough to play
>against on 3-ply, that will just cause confusion, but if I say 4-ply more
>people will understand what I mean. Still more people would understand
>exactly what is meant if gnu didn't use a different definition.
>There would be a clear-cut standard now if gnu were to follow the general
>usage, rather than trying to use a different definition.
I hear your concerns about standardizing terminology and ease of
communication. As a general principle, I agree with you. I've
disagreed with gnubg's definition of cubeless money equity with
the Jacoby rule, and I dislike its gammon prices, which do not
correspond to the way these have been calculated for twenty years
or so.
In this case, however, we have had competing definitions for some
time, so there is still room for choosing between the two as we
standardize. A number of us have written that we find the "count
from 0" more logical. We prefer it. We think: why not standardize
on this terminology? The fact that Zbot's code uses 0-ply for
static evaluations indicates that someone on your team at least
used to think that this labeling was more logical.
You envision gnubg changing to Snowie's terminology producing a
definition shared by gnubg, Snowie, and Zbot. JellyFish differs,
but it is less important because it has been superceded
by newer programs.
Let me imagine another scenario: gnubg continues to count from 0.
And Zbot counts this way, too. Now there is a definition shared
by JellyFish, gnubg, and Zbot. Snowie differs, but over time it
becomes less important as it is superceded by newer programs with
more features and better play, like Zbot and gnubg.
David Montgomery