From 1997 and onwards i managed to join in the computer chess world
champs every year.
Besides the participants Stefan MK (Shredder), Shay Bushinsky and
Amir Ban (both junior)
and tournament director Jaap van den Herik and Joke Hellemons who is
doing the entire
organisation from ICGA side; besides that 'hard core' addicted
computerchess guys and woman,
all of them excellent company also in restaurants, the other few who
show up are there at most
just for a year or 3.
If we forgive sometimes that people cannot show up 1 or 2 events,
then we can also add
Gerd Isenberg (who might be missing in Beijing) and David Levy.
Especially David is even better company in a restaurant, sometimes
he pays the bill.
Note that in ICGA events it is very easy to find the board (Jaap,
David and Joke).
In Paderborn 1999 someone needed them, so we adviced the guy to look
on the internet and search for the most expensive
hotel and most expensive restaurant; indeed in hotel Arosa the ICGA
board had dinner.
I do not blame authors to not show up at events. Showing up is
expensive, and most authors have many talents and get
asked for many different type of events. Most events you have to pay
for.
Most of them only show up when their engine is 'in big shape'. You
can be sure that the strongest engines also join an event.
An author knows pretty well when his engine makes a chance of doing
well. The same will be true for computer-go.
The programmers who know they make a chance of winning, they all will
be there for sure.
From the above elite group of ICGA members/participants that shows
up a lot at the computerchess, at least 1 is not bribable.
Maybe that's why i am looking for a job now. Even having no sureness
of income as of now,
it would not occur to me, not even at gunpoint, to recognize some
other event as the official world championship,
when it is clear that the entire organisation was officially planned
and scheduled for China.
It is not a problem when others organize events, in contradiction,
when others also organize events that's even better.
It is nice to have many events, that makes a science more popular.
Yet do not forget that the real and official event is in Beijing.
That aside, every year, when the location is far away, it is tougher
to participate;
the struggle to find a sponsor paying my bills is very tough.
Sometimes there is years i succeed in that,
other years i pay it myself.
The ticket costs i have are far larger than the amount of money ICGA
pays back. Add a hotel and extra travel time (as
the airplane to it needs like 10 hours or so, so you lose quickly 2
days). Additionally being born in a nation where it hardly
ever gets hot, i cannot stand heat very well. I figured that out in
Tel Aviv 2004.
In computerchess there is another cost, that most of you ignore.
Hardware. Easy if you come from Japan, USA, Germany, UK or France.
I come from a tiny nation. There is no highclocked 8 core machines
here that i can get from sponsors to test and play at in this tiny
nation of 16.5 million inhabitants.
Now that computer-go gets slowly more mature in this sense that
engines start to beat slowly more human beings,
and not soon from now will beat 99% of all go-players, after which it
will take endless until you can conclude objectively
that they are stronger. Of course for the average Joe on the street,
he will conclude 10 years before it happens that the
machine is stronger, as the go world will figure out within a year or
10 from now that the human side has a major weakness
that will cause it to lose 10 years sooner than it should: a human
being is in big need of money and therefore lose for money.
This is partly because of arrogance. They do not prepare and play
silly against the machine, usually several 10 kyu moves
get made during the game, to still let the game look interesting and
especially to give the sponsor an interesting
result. This instead of beating it bigtime. The stronger the player,
the more dumber social. It will be one of those players who will lose.
If the best player cannot get bribed, like Topalov refused a match
under the conditions offered,
they'll skip and wait until a next world champion is there who is
bribable (why do those always have a Russian passport?).
Even winning 10 matches after by human side, or losing games to tiny
Grandmasters, that doesn't matter. The first match a world top
player loses will get all attention. The damage will be done and the
average Ye on the street will consider the machine stronger.
Ye is always right; this is because Ye follows the first marketing
principle. The first marketing principle says:
"It is better to be the first than the best"
(Al Ries & Jack Trout)
For a titled player like me, comparable to about 1-4 dan professional
(i'm 4 dan only and only when a game is real important to my team,
in all others i'm at most 1), i needed a bunch of years to draw that
conclusion; it was rather hard to swallow.
Kasparov lost in 1997 from Deep Blue. A 10 ply searching chess engine
in the opening, forward pruning in the last few plies even.
Todays go programs search deeper than that, to give an example. In
the year 2000 when eating in a fastfood Hamburger restaurant,
i played a blindfolded match against a 10 ply searching program.
Though very bad and from France, i beated it blindfolded.
You get the picture.
Now in 2008 i'm pretty sure machine is stronger.
Not because it is anyhow qualitatively better than me, but sometimes
i make mistakes in those bizarre positions they play,
and then you're dead against a machine making a single big mistake.
Go is a game where that type of play is possible even better.
They create total dubious chaos at the board in a dubious manner, but
not dubious enough to not lose directly,
and then when being in that chaos, the machine will toast you when
the number of possibilities get too much.
Go is a far better game objectively to create such chaos.
Thiefery, cheating, bribing and losing for money works. In the 21st
century, Kasparov still got
match offers against the machine. Both for being moneygrabbers, may
he and Fischer rot in hell.
Sometimes there is some copycats from the middle east, who produce
some sort of a machine with FPGA cards which are
very low clocked.
Now we can of course first do a technical discussion. Namely that
showing up with low clocked fpga cards in 2004 around 30Mhz
is pretty pathetic. Even if you get 3-5 million nps, that's still
pathetic considering its design; it shares that crappy design with
deep blue.
Not ordering moves in a good manner and no caches loses you soon
factor 10. It only searched 18 ply worst case in a dubious manner.
Engines in 2005 single core to quadcore already got 18 ply.
An engine in hardware has huge if you either do one of the next 2
things:
a) incorporate huge amounts of knowledge
b) making a multicore hardware processor swallowing really a big surface
that is high clocked and has caches in hardware
Deep Blue and Hydra/Brutus failed in both respects.
The fast chessprograms (so the ones with tiny evaluation function,
and as far as i know my chessprogram Diep is the only one with a
really big
evaluation function), when programmed in C, so not even fully in
assembler, are around 1000 cycles a node.
So at a todays overclocked 8 core machine of 4Ghz, their only
limitation in speed is the memory controller.
Achieving a 10-20 million nps they easily do, just limited by the
memory controller.
The hardware guys who made chessprograms with near to no
chessknowledge, to move ordering and other tricks you soon need 10-20
cycles
a node. So a FPGA card at todays 60Mhz or so can deliver at most 6
million nps.
Now that would be a phenomenal speed if it was my chessprogram which
searches 200k nps at a 3Ghz core2,
yet that isn't the case. Especially for my program getting 6 million
nps in hardware isn't much.
Search is so so inefficient in hardware, what i need to do the last
few plies with Diep, to do that in hardware is not impossible,
but it eats a lot of sequential steps. Each sequential step in
hardware is a cycle. Soon you're 40 to 60 cycles to do that.
Even then, you still miss RAM to cache searches. Even a 1024KB
hashtable on-chip, would make a hardware chip fly.
This is the big mistake both Donninger and Hsu made.
They delivered both hell of an achievement perhaps in the eye of the
hardware layman, they failed algorithmically like total laymen.
Hydra is factor 50 slower clocked on each processor, 30-60Mhz versus
cpu's 3 to 4Ghz.
Deep Blue was around 30Mhz clocked when cpu's were 300Mhz. Hsu was
from hardware viewpoint much better than Chrilly,
making his CPU in VLSI, yet he for example even forgot to use
killertables in search. Something real simple and known
since the 80s to work well. Chrilly did include killertables but
forgot to create hashtables, total crucial in todays computer
chess search.
'forgetting' hashtables is really a beginnersmistake. No matter how
hard it is to make them.
If you do not know how to make a transpositiontable in hardware,
don't make a cpu.
A go/chess chip without transpositiontable is just a marketing chip.
It has no significance from technical viewpoint seen; it doesn't
speedup the search, using its nps as a marketing instrument is
about the only thing it gives.
Todays cpu's you buy in the store, also are as good as its L1, L2 and
even L3 caches are.
The funny thing about Hsu is that where he is a hardware genius
getting a chessprogram to work on a chip already in the 80s,
he just cares about search depth in his thesis. Deep Blue however was
a total marketing machine. Some oldie RS6000 technology,
combined with a hardware chess cpu that got 10/11 ply in opening and
later in game 12 ply. The machine had a theoretic capability
of searching half a billion nodes per second or so. The funny thing
is that just using nullmove, a technique well known to be a winner
technique by 1995, as Frans Morsch told everyone it kicked butt for
him and won him the world title (and he deserves credits for that),
Deep Blue 1997 still isn't using nullmove.
At a 1024 processor supercomputer, using the biggest partition of 512
processors, i did do several runs of Diep,
without using last few plies a hashtable. That really hurts,
especially when running parallel. The more cores and latency problems
you have communicating between the different search processes, the
more difficult it is to achieve a good branching factor.
Move ordering in software is real easy to do whereas in hardware and
at GPU's it is real real tough.
Hydra therefore lose a factor 10 or so of there speed the last few
plies. What Deep Blue lost there we cannot even estimate.
Both Hydra and Deep Blue are hard to see as hardware chessprograms.
They aren't doing entire search in hardware.
Their hardware concept is so ugly bad from branching factor viewpoint,
that both programmers decided to get the search into software as soon
as possible,
doing the utmost minimum of search in hardware.
Hydra is doing mostly 2 to 3 ply searches in hardware, Deep Blue was
doing 4 ply in hardware. Basically we must see the
cpu's as 'coprocessors' therefore, which only for like 1 year give
some small boost to what is then the latest CPU.
I see Hydra as a cluster of 64 processors 2.8Ghz P4 which was strong
end of 2004, start 2005.
Even then, a 64 processor cluster still is good hardware. I'd be
happy with a 64 processor cluster.
It achieved about 16-20 ply searches in the events i played it.
Todays programs at 4 cores, if i look to the world top, are also
getting worst case about 18-20 ply first few moves out of book,
quite a lot more after that (hydra doesn't scale there nor did deep
blue, showing their hashtable weakness to full extend).
Knowing it forward prunes in hardware last few plies and is doing
something dubious in software search,
a hardware cpu of course only makes sense to produce when you clock
it to 300Mhz, have 32 or more cores,
and have for each core its own hashtable, at least 512KB. So you do
everything at 1 cpu.
Additionally you need more chessknowledge than todays fast programs;
they objectively play real real passive chess,
which wins for them because playing agressive chess (and IMHO
attacking is objectively the best way to play chess),
requires a lot of chessknowledge.
Such a cpu will need at least 30 cycles to implement all algorithmic
knowledge and sequential enhancements,
as well as that huge evaluation function.
So the entire search would be inside 1 cpu, with as most important
feature of the chip having a hashtable in hardware
of at least 512KB. That gives a 300 million nps monster. It would be
a real big CPU by the way and only using 1 makes
sense. Putting several of them in a cluster doesn't make sense. You
want the entire search in hardware.
If you're not doing that, then it is better to not make a cpu at all,
except from marketing viewpoint.
Chess and also Go at those search speeds are just too dependant upon
hashtable to not use it. My experiments using 460 processors
and giving each processor just 1MB of hashtable indicated that even
such a tiny hashtable still performed very well. In fact it just lost
1 ply over using a 200 fold bigger hashtable, this at a single
position of 10 hours in total.
Would a go-cpu also be useful not using a transpositiontable?
The answer is NO.
Is it complicated to make such a hardware go-cpu?
The answer is YES.
A SHARED transpositiontable (shared over all cpu's) in hardware has 3
important effects when having a lot of cores
searching of a hardware cpu (say 4+):
a) you avoid really a lot of futile parallel searches to get
accomplished, with tens of processors or even more,
this is really important
b) transposition cutoffs give a branching factor improvement giving
an exponential speedup
c) storing the best move and using this in the search really improves
the move ordering a lot,
in fact it cloacks the bad move ordering that a hardware design
has a lot; it gives a huge branching
factor improvement, far more than we get in software from it
and therefore a huge exponential speedup,
Just having a transpositiontable of 512KB to 1MB in a hardware cpu,
so ON chip, gives a hardware chess/go-cpu
nearly the same branching factor like you can achieve in software.
AFAIK no one ever has achieved this so far.
You can argue that hydra gets 220 million nps or so. Deep Blue on
paper got a 100 million nps or so, but that's just
paper. At least from Chrilly we know he's not lying too much about
his nps, maybe.
With Deep Blue i've heard too many numbers and their search
depth was just too little to realistically put more than just a few
percent of all 480 cpu's to work;
if you get 10 ply in total from which 4 is hardware, you have to
split at a depth of 6 ply all cpu's,
it is very tough to get 480 hardware chips to work there. The "good
branching factor", of just above 4, from deep blue
moving from 10 to 12 ply, is just because it can put more and more
processors to work at 11 and 12 ply.
Yet you can argue that Hydras 220 mln nps got used ugly bad.
DeepSjeng at 4 million nps is getting the same search
depth at a 2.4Ghz quadcore (overclocked a tad to nearly 3Ghz). I know
DeepSjeng's search is comparable in dubiousity to Hydra's.
So somewhere Hydra loses a factor 50 search efficiency.
Note that this is not uncommon.
Deep Blue lost way more than that.
That is not a surprise to me. Making a program run well on a
combination of different architectures (hardware fpga cpu and at a
cluster),
is very difficult. He had to do it all himself. And as usual when you
do not visit ICGA events and do not talk in restaurants/pubs with other
chessprogrammers at the many other computer chess events, you soon
are not up to date with the latest search algorithms and tricks.
Of course once you choose for a certain setup, knowing it takes a
year or 5 to mature the hardware and software, the scene has probably
changed algorithmically bigtime by then. It is not easy then to just
change the entire search concept i guess.
However all this is not so relevant. We will not see Hydra in any
event because of the second marketing principle:
"If you cannot be first in a category, then create a new category
where you can be first"
(Al Ries & Jack Trout)
The category where Chrilly and the sheikh picked to be first was:
"the unbeatable chess machine".
That means of course they can join never an event where other
chessprogrammers have good hardware.
The risk of losing a number of games and not winning the title in an
'unbeatable manner' is too big there.
In fact i'd argue it is 99%, if i take the ceiling.
That type of event is called world championship.
Sometimes people show up at supercomputers, usually the programmers
from big nations have some great high clocked machine,
and nearly all programs are in big shape and kick butt at such event.
So the odds of winning that title is not so big for a chess machine
that gets outsearched by other opponents and has less chessknowledge
than them.
Computer go right now is at a phase, where the programmers who slowly
learned how to make an evaluation function, now get kicked butt
by better search algorithms. Also not too many cores get thrown into
action yet.
So i'd argue that the time is there that some sort of hardware/
software professional should show up in computer-go,
paid by some rich sponsor who wants to have a shot at being the first
beating a strong professional player.
I'd argue it is best to talk to Chrilly for that.
For such a sponsor to get a kick butt go-machine accomplished,
Chrilly is the best choice by far. Algorithmically he's good enough
to not be
that much behind the best and he has learned what to not do in
hardware. Maybe next machine doesn't lose a factor 50 somewhere.
From my viewpoint by far most important advantage of Chrilly is:
he's not keeping his mouth shut about what he algorithmically is doing.
For science this is very important. Guys like Bushinsky and Meyer-
Kahlen tell nothing about what they do inside their engine.
Chrilly is a honest person there.
Means his go playing machine might kick major butt for a short while.
Of course after his machine has a reputation,
he'll not show up of course at computer-go events anymore and he'll
spend a few years playing different professional go
players whose go playing capabilities will get dramatically
lobotomized by the bad oversight at the board, caused by the dollar
signs in the eyes.
However when such professionals show up in the ICGA Olympiads, then
you'll figure out ICGA is a genius organisation.
Though all amateurs spit at it, as they see person A get paid and
person B doesn't which they find unfair, meanwhile the ICGA themselves
grabbing even more (hint: try to figure out what they charge as an
'organisation fee'). All that doesn't matter. Everything is negotiable
and during the event the ICGA has the one person that no other has.
ICGA has Jaap van den Herik as tournament director.
He is a genius in getting rivalling commercial professionals at 1
table and keep the peace.
Vincent
On Jul 2, 2008, at 9:12 PM, Ian Osgood wrote:
On Jul 2, 2008, at 10:31 AM, Zach Wegner wrote:
On Tue, Jul 1, 2008 at 4:33 PM, Erik van der Werf
<[EMAIL PROTECTED]> wrote:
That's a pretty good deal!!!
http://64.68.157.89/forum/viewtopic.php?
topic_view=threads&p=193819&t=21591
Why isn't there any sponsoring like this for the other tournaments?
Erik
They pretty much have to. The ICGA has achieved a rather lousy
reputation in the chess community, and very few participants are
showing up nowadays (11 on the list now). Compare that to the online
tournaments, which always have around 30 or more participants.
Personally I'd like to go and meet some other programmers, but there
are so few. And even after the subsidies it would still be very
expensive...
In my opinion, the size of the ICGA World Computer Chess
Championship event is irrelevant. More importantly, it has the
prestige to consistently draw the top candidate programs. This
year, past champions Rybka, Junior, Shredder, and HIARCS will be
competing for the title. (The author of Zappa no longer develops
his program, so I'm not surprised he dropped out. I don't know why
Fritz never attends. Hydra would also be an interesting participant
as one of the last custom chess supercomputers.)
By contrast, the ICGA Go events never get top candidate program
participation, and before this year have had smaller turnouts than
the chess event. Since the expiration of the Ing Prize, the last
event of any kind which had such participation was the 2003 Gifu
Challenge (KCC Igo, Haruka, Go++, Goemate, Many Faces, GNU Go, Go
Intellect, Aya, Katsunari). The size of this year's event is
encouraging, but where are Go++, Haruka, HandTalk, and GNU Go? And
what ever happened to Wulu and GoAhead?
Ian
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
