TD-Gammon, NN, AI, human bias, etc.

[mu...@compuplus.net]

On January 12, 2020 at 11:49:35 AM UTC-7, Tim Chow wrote:

> On January 12, 2020 at 10:59:22 AM UTC-5, bgbl...@googlemail.com wrote:

>> Am 11. Januar 2020 19:20:33 UTC+1 schrieb Tim Chow:

>>> Does BGBlitz include "human biases" or does it learn
>>> purely from self-play?

>> It depends on what you regard as "human bias".
>> The inputs are more or less in line with what Berliner
>> already did, so are expert inputs "Human bias"?

Of course. What else?

>> BGBlitz learns purely due to self play.

> Well, I assume that what Murat is complaining about is
> the distinction between versions 0.0 and 1.0 of TD-Gammon,
> as described for example here:

Why do you characterize my arguments as "complaining"?

I'm pointing out defects in so-called "NN" or "AI" bots
to show that they are not "NN" or "AI" by definition.

See this article:

https://www.cs.cornell.edu/boom/2001sp/Tsinteris/gammon.htm

that talks about creating a bg bot "free from the biases
of existing human knowledge".

If I follow the versions correctly, first TD-Gammon was
trained unsuperwised, through self-learning without any
human bias and could only play cubeless single games.

Tesauros first bot Neurogammon had relied on supervised
learning as did TD-Gammon starting with the second version.
Human bias was added back to make it play cubeful matches.

Here is another article by Tesauro:

https://www.bkgm.com/articles/tesauro/tdl.html

where he says:

"Strategy for use of the doubling cube was not
"included in TD-Gammon's training. Instead, a
"doubling algorithm was added after training
"that makes decisions by feeding TD-Gammon's
"expected reward estimates into a theoretical
"doubling formula developed in the 1970s.

How is that for "AI"?

I say it's "FAI", "Fartificial Intelligence"...

> http://papers.nips.cc/paper/1302-on-line-policy-improvement-using-monte-carlo-search.pdf

Apparently this version was cubeless also since he says:

"In future work, we plan to augment the program
"with a similar Monte-Carlo algorithm for making
"doubling decisions.

What is also interesting in that article is about the CPU
power, where he says:

"Our Monte-Carlo simulations were performed on the
"IBM SP1 and SP2 parallelRISC supercomputers at IBM
"Watson and at Argonne National Laboratories. Each
"SP node is equivalent to a fast RSj6000, with
"floating-point capability on the order of 100 Mflops.
"Typical runs were on configurations of 16-32 SP nodes,
"with parallel speedup efficiencies on the order of 90%.

Let's say 24 nodes x 100 Mflops x 90% = 2.16 Gigaflops.

Today we have $2,000 desktop PC's achieving Teraflops.

IBM's latest supercomputers are pushing 200 Petaflops,
with even the Blue Gene/Q (or whatever the latest) is
capable of 20 petaflops. That's 10,000,000,000 times
the power of 24 SP2's...!

With that, now you all should re-read all those articles
striking out everything related to CPU power limitations
and/or costs, which was their excuse of taking shortcuts
to substitute human bias for machine self-learning match
play and cube play.

BTW: I just searched Tesauro and found this:

https://researcher.watson.ibm.com/researcher/view.php?person=us-gtesauro

Apparently hes is still quite young (60ish?) and is still
at IBM. Considering his 30+ more years of seniority since
TD-Gammon, I'm sure sure IBM would spare a couple of Blue
Genes for him to work on a AlphaZero BG bot...

MK

[Tim Chow]

On Tuesday, January 14, 2020 at 7:00:30 PM UTC-5, mu...@compuplus.net wrote:

> If I follow the versions correctly, first TD-Gammon was
> trained unsuperwised, through self-learning without any
> human bias and could only play cubeless single games.

This is my understanding also.

> Tesauros first bot Neurogammon had relied on supervised
> learning as did TD-Gammon starting with the second version.
> Human bias was added back to make it play cubeful matches.

I don't think that this is quite accurate.

There are (at least) two ways to add "human bias." One is to use
hand-crafted features in one's description of a board position (as
opposed to just specifying the number and color of checkers on each
point). The second way is to introduce human expert judgment when
training the network. These two adjustments are very different from
each other.

Certainly, TD-Gammon introduced hand-crafted features in later versions.
I'm less clear on whether it introduced human expert judgment during
the training process. The mention of an "expert data set" seems to
suggest that it did, but so far I've found only rather vague statements
about this that leave me somewhat uncertain.

Current bots don't use human expert judgment when training the network.
(This is what Frank Berger meant when he said that BGBlitz uses self-play
only.) But I think they all use more than the raw board position as input
to the network.

I also don't believe that the *reason* for introducing human bias was
to make it play cubeful matches. It was to improve the performance. I'm
assuming that they tried playing version 0.0 against 1.0 and found that
version 1.0 had a persistent edge against version 0.0.

Making the bot play cubeful matches was just a matter of plugging in a
MET and using a Janowski formula for cube actions. Whatever you think of
these techniques, they have nothing to do with the setting up and training
of the neural network itself. The net is still trained cubelessly, and
the cube decisions and match play features are added on at the end, treating
the network as a "black box."

---
Tim Chow

[mu...@compuplus.net]

On January 14, 2020 at 6:35:03 PM UTC-7, Tim Chow wrote:

> I'm less clear on whether it introduced human expert
> judgment during the training process. The mention of
> an "expert data set" seems to suggest that it did,
> but so far I've found only rather vague statements

I have also seen the expressions "unsupervised trained"
for the initial and "supervised trained" for the later
versions. Why else could they have been used other than
to mean human bias was introduced.

In addition, I wasn't suggesting that human bias could
only be introduced during training.

> Making the bot play cubeful matches was just a matter
> of plugging in a MET and using a Janowski formula for
> cube actions.

That's what I call "human bias".

> Whatever you think of these techniques, they have
> nothing to do with the setting up and training of
> the neural network itself. The net is still trained
> cubelessly, and the cube decisions and match play
> features are added on at the end

Yes, I meant the same thing. The miscommunication may be
in that I also said that the bot should learn match play
and cube decisions through training (self-play) instead.

I'm not opposing MET's but sayin that separate MET's for
cubeless and cubeful matches should be built by the bot,
not by humans based on human play.

Similarly, the bot should learn how to make the best cube
decisions through self-play, not from Jack, Janowski, etc.

I predict that a true NN/AI bot will improve significantly
in adjusting its checker playing according to match length
and score.

I also predict that such a bot will make shockingly different
cube decisions than the current bots.

MK

[Tim Chow]

On Wednesday, January 15, 2020 at 7:45:10 PM UTC-5, mu...@compuplus.net wrote:
> I predict that a true NN/AI bot will improve significantly
> in adjusting its checker playing according to match length
> and score.
>
> I also predict that such a bot will make shockingly different
> cube decisions than the current bots.

This is an interesting prediction. I think that eventually someone
is going to create "AlphaGammon" and we'll find out the answer. It's
getting easier and easier to do this sort of thing, so eventually
someone may create "AlphaGammon" for a master's thesis or even an
undergraduate thesis.

I certainly believe that AlphaGammon will make dramatically different
decisions from current bots in *some* positions, such as superbackgames.
But I don't expect it to make significantly different decisions in most
positions.

I wonder if you have any specific positions in mind? Can you devise
some positions where you predict that AlphaGammon will make a "shockingly
different" cube decision from the current bots?

---
Tim Chow

[Paul Epstein]

On Thursday, January 16, 2020 at 1:00:00 AM UTC, Tim Chow wrote:
> On Wednesday, January 15, 2020 at 7:45:10 PM UTC-5, mu...@compuplus.net wrote:
> > I predict that a true NN/AI bot will improve significantly
> > in adjusting its checker playing according to match length
> > and score.
> >
> > I also predict that such a bot will make shockingly different
> > cube decisions than the current bots.
>
> This is an interesting prediction. I think that eventually someone
> is going to create "AlphaGammon" and we'll find out the answer. It's
> getting easier and easier to do this sort of thing, so eventually
> someone may create "AlphaGammon" for a master's thesis or even an
> undergraduate thesis.
>
> I certainly believe that AlphaGammon will make dramatically different
> decisions from current bots in *some* positions, such as superbackgames.
> But I don't expect it to make significantly different decisions in most
> positions.

...

Your (Tim's) expectation makes sense (to me) when I compare the situation
to chess. Chess and backgammon are the only traditional board games of
which I have considerable knowledge.
If you compare the chess played by AlphaZero with that of the strongest
human players, the overwhelming majority of AlphaZero's moves are moves that
are either a) typical of human experts or b) moves that human experts
would play if they had a day to do the thinking rather than just two
minutes or c) moves that could be made by human reasoning if humans were
able to examine more positions.
It's not as if AlphaZero is likely to open with 1. h4, and it's not as if
it regards a bishop as stronger than a rook.
All the traditional theory about strong pawn structures etc. held solid.
The computer revolution in chess (before machine learning though) did
bring a few surprises but surprisingly few.
Some of these surprises are:
1) If the rook starts out in a good position, queen and king vs rook and
king is an extremely difficult theoretical win (for humans) and requires
extensive study. If a computer has the rook starting from a good position
and if the opponent hasn't spent at least several hours learning this
endgame, the human has no chance of a win. Previously, this had been
thought to be a completely trivial endgame.

2) Rook and bishop against rook and king is better for the rook and bishop
than previously thought. There are more winning positions for the rook
and bishop than was thought, and also many of the drawing positions
only enable drawing against a computer with extensive study.

Paul

[Tim Chow]

On Thursday, January 16, 2020 at 5:11:23 AM UTC-5, Paul Epstein wrote:
> The computer revolution in chess (before machine learning though) did
> bring a few surprises but surprisingly few.

I think that another surprise was that king and queen versus king and
two knights was, in the old days, thought to be usually a draw. But
now we know that it is almost always a win for the queen.

I think also that the chess engines underscored how much the evaluation
of a position depends on concrete calculation rather than general
strategic principles. This was something that the top players were
starting to grasp in the late 20th century but which really became
apparent when the engines came into their own. A lot of positions that
the "old masters" would have quickly evaluated as (say) "lost for Black"
have turned out to be equal or even better for Black. Defensive
technique has dramatically improved with the advent of the engines,
and certain things that used to be regarded as huge positional advantages
or disadvantages have been found to be not as decisive as formerly believed.

---
Tim Chow

[Philippe Michel]

On 2020-01-15, Tim Chow <tchow...@yahoo.com> wrote:

> There are (at least) two ways to add "human bias." One is to use
> hand-crafted features in one's description of a board position (as
> opposed to just specifying the number and color of checkers on each
> point).

If you consider the nature of the game, where a blot and a point have
very different properties, it is using the number of checkers that
introduces a human bias and (the most basic of) the representations used
by Tesauro that are natural : is the place empty ? is it a blot ? is it
a point ?

[Paul Epstein]

I'm not sure what you mean by "almost always" in the context of a queen
beating two knights. In normal chess parlance, this is considered an
abbreviation for "In almost all non-trivial positions".
The great majority of positions are trivial and include situations where
the queen immediately captures or forks or pins one of the knight.

With this usual interpretation of "almost always", you've made a false
statement, perhaps by misremembering. I knew this even before I checked by
googling. A good mnemonic for remembering the fact that Q v 2N is actually
a draw is John Nunn's remark that this is the only drawn material balance
which stays a draw when one of the sides loses a queen for nothing.

Humans generally bias their language and opinions to make their activities
seem nobler and more sophisticated than they are. Most people view chess
calculation as cruder than positional judgment so players will tend to
exaggerate (until proved otherwise) the relative importance of intuitive
positional judgment vs calculation.

Paul

[Tim Chow]

On Saturday, January 18, 2020 at 4:15:20 PM UTC-5, Paul Epstein wrote:
> With this usual interpretation of "almost always", you've made a false
> statement, perhaps by misremembering. I knew this even before I checked by
> googling. A good mnemonic for remembering the fact that Q v 2N is actually
> a draw is John Nunn's remark that this is the only drawn material balance
> which stays a draw when one of the sides loses a queen for nothing.

You're right that I misremembered. I remembered a figure of 90%+, but as
you say that includes a lot of trivial positions.

Still, I do believe that the theory changed significantly with the advent
of tablebases. Suppose we take the positions where the side with the
knights moves first. Then according to this source---

https://lichess.org/forum/general-chess-discussion/endgame-queen-vs-2-knights2-bishops

---the side with the queen wins the majority of the time (though not "almost
all" the time). I believe that this was a surprise.

Queen versus two bishops was, I think, also thought to be usually a draw,
although I think that it was understood that the two knights provided
better defense than the two bishops on average.

---
Tim Chow

[bgbl...@googlemail.com]

Am Mittwoch, 15. Januar 2020 01:00:30 UTC+1 schrieb mu...@compuplus.net:

> I'm pointing out defects in so-called "NN" or "AI" bots
> to show that they are not "NN" or "AI" by definition.

AFAIK there is no common definition was intelligence is, so I regard it as even more difficult to define what real AI is. There are two common definitions that of a real AI (what nobody knows what is is) and weak AI, i.e simply put expert in one area and an idiot in all other.

Further I think that e.g. chess programs are called Ai, although their abilities (before Alpha chess) were programmed algorithm, maybe tuned somehow but plain algorithm no leaning at all.

BGBlitz learns purely by self play the winning probabilities of cubeless money games.

> https://www.cs.cornell.edu/boom/2001sp/Tsinteris/gammon.htm
In fact this article is not very interesting. Unless I overlooked something you a recreation what Tesouro published.

> that talks about creating a bg bot "free from the biases
> of existing human knowledge".

I believe that your opinion, that the so called expert input introduce what you call "human bias" is simply wrong:
- first also called "expert inputs" most are pretty simple evaluations, like pip count, probability to hit, probability to come in from the bar etc. I can't see that this introduce a bias. Further if you introduce some weird input, let's says the time needed for the move, the color of the board, anything really strange, the net will find out on its own that it is non sense and ignore it sooner or later.
- a couple of inputs can be easily developed by the net itself, so why use them? The effect of things like pip count is only to speed up learning. If you look at the article they trained the net with 50.000 games. There you need any speedup you get. On a single core of an old i7-2600 BGBlitz plays about 2 million games per day! (IIRC, I'm too lazy to boot my training computer). Most probably the net will develop this concepts on is own.

> "Strategy for use of the doubling cube was not
> "included in TD-Gammon's training. Instead, a
> "doubling algorithm was added after training
> "that makes decisions by feeding TD-Gammon's
> "expected reward estimates into a theoretical
> "doubling formula developed in the 1970s.
> How is that for "AI"?

I guess you mix up the concepts of "learning by itself" and AI. As mentioned above, chess programs were AI but didn't learn by itself (in earlier years). BGBlitz learns by itself playing cubeless money game and the rest is algorithms. It might be an interesting experiment to train for cubeful evaluations, maybe I'll do it if I have a little more time.

By the definitions I'm used to this is AI, but I wont claim more than that BGBlitz plays pretty good backgammon. At least I'm very confident that BGBlitz will have an positive equity against any human in match play on the long run.
You may have different criteria of what you regard as AI, that's completely o.k. albeit I assume my point of view is more common.

> Apparently hes is still quite young (60ish?) and is still
> at IBM. Considering his 30+ more years of seniority since
> TD-Gammon, I'm sure sure IBM would spare a couple of Blue
> Genes for him to work on a AlphaZero BG bot...

I'm pretty sure that AlphaZero wont be a breakthrough because that was what Tesauro already achieved. There might be some incremental achievements, partly because Google can throw a lot of HW on any problem (fun fact: learned better to play than Stockfish in less than a day seems to prove huge supremacy of Google and/or what a simple game chess must be. If you scale that to the HW any plain programmer is able to buy you are around a year of training).

[Tim Chow]

On Thursday, January 23, 2020 at 5:24:26 PM UTC-5, bgbl...@googlemail.com wrote:
> I'm pretty sure that AlphaZero wont be a breakthrough because that was what
> Tesauro already achieved. There might be some incremental achievements

This depends on how you define "incremental." For me, being able to accurately
play a backgame with 15 checkers back would be a major improvement, not merely
an "incremental" one, since current bots completely misunderstand such
positions.

---
Tim Chow

[bgbl...@googlemail.com]

Am Freitag, 24. Januar 2020 01:14:36 UTC+1 schrieb Tim Chow:

> This depends on how you define "incremental." For me, being able to accurately
> play a backgame with 15 checkers back would be a major improvement, not merely
> an "incremental" one, since current bots completely misunderstand such
> positions.

Just played two games of Snake, one from either side. To me it feels not like "completely misunderstand".
BTW according to Wachtel Snowie 4 might have handled this reasonably too, but I can't check that.

[Tim Chow]

On Friday, January 24, 2020 at 5:53:14 PM UTC-5, bgbl...@googlemail.com wrote:

> Just played two games of Snake, one from either side. To me it feels not
> like "completely misunderstand".

Start with this position: X has 2 checkers on each point of O's home board
and 3 checkers on O's bar point. O has a single checker somewhere, and
14 checkers borne off. Are you telling me that the bot correctly rolls the
prime around the board? And its cube action seems right to you?

---
Tim Chow

[mu...@compuplus.net]

On January 23, 2020 at 3:24:26 PM UTC-7, bgbl...@googlemail.com wrote:

> Am 15. Januar 2020 01:00:30 UTC+1 schrieb mu...@compuplus.net:

>> I'm pointing out defects in so-called "NN" or "AI" bots
>> to show that they are not "NN" or "AI" by definition.

> You may have different criteria of what you regard as AI,
> that's completely o.k. albeit I assume my point of view
> is more common.

Fair enough. Everyone got to clarify what they meant by AI.
As long as we know how the bots actually "learned" to play,
there is no need to argue which definition of AI is better.

> I'm pretty sure that AlphaZero wont be a breakthrough
> because that was what Tesauro already achieved.

Not true at all! What Tesauro had achieved was the same as
what you have achieved.

Like the AlphaChess learning to play without opening books,
AlphaGammon will learn to play without Jackoffski's formula!
And without human compiled MET's! And without whatever other
"human bias" hard-coded into it!

> There might be some incremental achievements, partly because

> ... If you scale that to the HW any plain programmer is able
> to buy you are around a year of training.

If from AlphaGammon you understand a bot only capable of
jacking off faster but still using Janowski's formula, I
can see why your expectations only "incremental"...

MK

[bgbl...@googlemail.com]

Am Samstag, 25. Januar 2020 00:48:36 UTC+1 schrieb Tim Chow:

> Start with this position: X has 2 checkers on each point of O's home board
> and 3 checkers on O's bar point. O has a single checker somewhere, and
> 14 checkers borne off. Are you telling me that the bot correctly rolls the
> prime around the board? And its cube action seems right to you?

Just played a couple of games. I have to eat my words (at least partially). BGBlitz tries to get a prime on it's half of the board, but it doesn't see that the prime could be simply rolled through the outfield. Albeit that wouldn't match my definition of "completely misunderstand" BGBlitz surely doesn't get the point of rolling it home.
I just checked my code. In fact one of my "expert inputs" only checks the existence of a prime from the opponent bar on. Although this position provbably has never been reached in a real game, I'll think about what I can do ,,,,

[Axel Reichert]

bgbl...@googlemail.com writes:

> Am Samstag, 25. Januar 2020 00:48:36 UTC+1 schrieb Tim Chow:
>
>> Start with this position: X has 2 checkers on each point of O's home board
>> and 3 checkers on O's bar point. O has a single checker somewhere, and
>> 14 checkers borne off. Are you telling me that the bot correctly rolls the
>> prime around the board? And its cube action seems right to you?
>
> Just played a couple of games. I have to eat my words (at least
> partially). BGBlitz tries to get a prime on it's half of the board,
> but it doesn't see that the prime could be simply rolled through the
> outfield. Albeit that wouldn't match my definition of "completely
> misunderstand" BGBlitz surely doesn't get the point of rolling it
> home.

An even tougher test for the bots (I played around with GNU Backgammon)
seems to be to use *two* checkers instead of one for O, say, on X's
points 1 and 2 in Tim's position. An example (advanced prime, to save me
the hassle from rolling it home) would be:

GNU Backgammon Position ID: AABAAbZtOwAAAA
Match ID : UQkAAAAAAAAA
+13-14-15-16-17-18------19-20-21-22-23-24-+ O: gnubg
| X | | | OOO 0 points
| X | | | OOO
| X | | | OOO
| | | | OO
| | | | OO
v| |BAR| |
| | | |
| | | |
| | | |
| X X X X X X | | | On roll
| X X X X X X | | O O | 0 points
+12-11-10--9--8--7-------6--5--4--3--2--1-+ X: axel (Cube: 2)
Pip counts: O 47, X 153

Quite frequently (when I played X and GNU Backgammon played O), O ended
up with an anchor at the edge of my full prime (I *was* able to roll it
home). From then on, the play is very difficult for X: Because O will
not be able to move, X has to decide whether to clear the prime from the
rear or to try a trap play.

Once O has escaped with hopefully only one checker, X will re-establish
his prime, continue rolling it home and try to pick up O's second
checker, which will eventually end up behind X's prime. The opportunity
for a trap play hence comes up repeatedly in case O gets an anchor at
the edge again.

In these cases, GNU Backgammon's evaluation of X's position and a full
roll-out differ considerably. Playing this position a couple of times
was very interesting for me (and also resulted in some nice positions
for testing my Isight count with race doubles).

Best regards

Axel

[Paul Epstein]

I have quite strong knowledge on these issues (not that I'm saying that
you don't).
Your statistic about the knights moving first does nothing to contradict
my point that Q v 2N is actually a draw in the usual sense, where trivial
positions are excluded. Even when we restrict our attention to the knights
moving first, we will usually get trivial wins for the queen because
the knights are likely to be far from their own king, and there will
be trivial and uninteresting wins for the queen involving 2 to 3 move combos
with attacking, pinning and forking. However, in almost all non-trivial
positions, this is a draw. A reasonable definition of non-trivial could be
that no side can win a piece or checkmate in <= 4 moves.
I don't see how this can be a surprise. A central queen attacks 27 squares.
A corner queen attacks 21 squares. Let's take the average and assume the
queen attacks 24 squares. Probabilistically, one of these squares contains
a knight. Then the question becomes: "Is the knight likely to have a move
escaping all sequences like check, check, fork allowing the queen to pick up
a knight?" The answer is no.
I'm not sure what you mean by saying that the two knights provide
better defence because there's such a huge difference in outcomes between
so many plausible sets of positions:
1) The set of all legal positions.
2) The set of all non-trivial legal positions.
3) The set of all positions which are in databases consisting of
actual games.
4) The set of all non-trivial positions in such databases.
5) The set of all legal positions which look plausible from a game.
6) The set of all non-trivial legal positions which look plausible from a
game.

I'll try to be a bit concrete.
Suppose it's the case that queen vs two knights and queen vs two bishops
occur with equal probability (and that's quite likely to be approximately
true).
Suppose I'm given a superpower where, if my games reach this stage,
I can use tablebases and computers to produce optimal play, but I can only
do this for one of the two types of endgame. Do I want the superpower
for queen vs two bishops or for queen vs two knights?
I'm taking the superpower for queen vs two bishops because I expect so
many of queen vs two knight positions to be trivial. So I don't see the
two knights as providing "better defence".

Paul

[Tim Chow]

On Saturday, January 25, 2020 at 6:31:09 PM UTC-5, bgbl...@googlemail.com wrote:
> I just checked my code. In fact one of my "expert inputs" only checks the
> existence of a prime from the opponent bar on. Although this position
> provbably has never been reached in a real game, I'll think about what I
> can do ,,,,

I think that there's a section in Woolsey's "Backgammon Encyclopedia" where
he notes that the "wheels come off" Snowie when the six-prime hits the
midpoint or so.

There was an actual tournament game between Mochy and another Japanese player
a few years ago, where Mochy's opponent went for an all-out superbackgame.
I don't know if I can find the game, but maybe someone else remembers. It
might be interesting to see what BGBlitz thinks of that game. Of course, in
complex positions, it's hard to be sure what the truth is. That's why I
focused on a simple position where it's clear that humans can play far better
than the bot can.

---
Tim Chow

[Tim Chow]

On Sunday, January 26, 2020 at 7:54:58 AM UTC-5, Paul Epstein wrote:
> I'm taking the superpower for queen vs two bishops because I expect so
> many of queen vs two knight positions to be trivial. So I don't see the
> two knights as providing "better defence".

Actually that's what I meant by "better defence" but I agree that the
term "better defence" was a poor choice of words. I meant that the knights
often let you set up a fortress, whereas there's no comparable simple
strategy for the bishops.

But back to the main question---most of the endgame knowledge that I have
stored in my head, I picked up from Paul Keres's "Practical Chess Endings."
When tablebases came out, I went through a lot of that book looking for
errors and found only a few. The Q vs. N+N and Q. vs. B+B positions were
fine, but what I seem to remember is that the tablebases revealed that the
Q won considerably more often than previously thought (where I'm using
Keres as my standard of "what was previously thought"). I realize that
"considerably more often" is a vague term, but are you telling me that this
isn't true? That the tablebases more or less validated existing knowledge
and there weren't any conceptual surprises in Q vs. N+N or Q vs. B+B, akin
to the surprises in Q vs. R?

---
Tim Chow

[Paul Epstein]

No, I think you're extrapolating too much. Prior to this post, you
posted twice on chess (in response to me).
In both of those posts, you made assertions about things being
"surprising". I don't think that your particular assertions about
which things were "surprising" are correct.
However, you can't deduce from that that I believe that there were no
"conceptual surprises" at all in the two endgames you mention.

With Queen versus two knights in random legal positions, it has always
been unsurprising that the Queen has plenty of forcing combinations so
the queen is likely to win even if you give the knights the move.
The surprise is that, when you restrict attention to the small percentage
of interesting positions where there isn't a simple 4-move-or-less winning
sequence, the result is actually (almost always) a draw.

So, re Q vs NN, the surprise is in exactly the opposite direction to the
direction than you think.
[However, I can't be 100% sure of my assertions because it's a long time
since I read up on this.]

Re queen vs two bishops, in interesting positions, you are absolutely
correct about the surprise -- the queen (almost) always wins and that is
a counter-intuitive result.

Paul

[Tim Chow]

On Monday, January 27, 2020 at 5:49:09 AM UTC-5, Paul Epstein wrote:
> So, re Q vs NN, the surprise is in exactly the opposite direction to the
> direction than you think.
> [However, I can't be 100% sure of my assertions because it's a long time
> since I read up on this.]
>
> Re queen vs two bishops, in interesting positions, you are absolutely
> correct about the surprise -- the queen (almost) always wins and that is
> a counter-intuitive result.

O.K., thanks for the clarification.

Last night, I took a quick look at Keres's book. I don't have it in front
of me right now, but my takeaway was that, with the caveat that the endings
were not completely "solved," his verdict for Q v N+N was "usually a draw"
and his verdict for Q v B+B was "it depends". (Implicitly, I think we can
assume that Keres was excluding "trivial" positions with an easy-to-see
immediate win for either side.) So if I understand you correctly, it was
surprising that the Q almost always wins against B+B. On the other hand,
it sounds to me that Keres at least wouldn't have been surprised that Q v
N+N is almost always a draw.

---
Tim Chow

[Paul Epstein]

I think Keres would also have been surprised by the strength of the queen.
I read the Keres reference. It's confusing because
it also says: "Black must of course post his pieces correctly."
That sounds a little bit like saying "I think you have the potential to
earn a lot of money but only if you find a high-paying job."

Then when you look at the diagram, he gives a standard fortress position.
In other words, he's saying that it's a draw if the knights achieve a
fortress. However, the surprise (even to Keres) is that the knights can often draw without beingable to achieve one of the fortress positions.

Paul

[mu...@compuplus.net]

On January 25, 2020 at 4:31:09 PM UTC-7, bgbl...@googlemail.com wrote:

> I just checked my code. In fact one of my
> "expert inputs" only checks the existence
> of a prime from the opponent bar on.

So, it's not the bot that wasn't able to figure
out how to roll a prime around but it was never
even given a chance and instead was blindfolded
by human bias...

> I'll think about what I can do

Why don't you let the bot think for a change..?

MK

[mu...@compuplus.net]

On January 26, 2020 at 11:36:06 AM UTC-7, Tim Chow wrote:

> Of course, in complex positions, it's hard
> to be sure what the truth is.

This is where we can learn the "statistical
truth" from self-learning bots but it may
not be "statistically" all that useful in
terms of human lifespan...

> That's why I focused on a simple position
> where it's clear that humans can play far
> better than the bot can.

You have this general problem with trying to
understand (and trying to explain to others)
complex subjects, using simple examples and
then extrapolating from them... :(

MK