Re: Turing test - Deep Blue

[Vincent Diepeveen <diep@xxxxxxxxx>]

At 05:00 PM 5/26/98 -0700, birk@xxxxxxxxxxxxxxxxx wrote:
>
>Vincent wrote:
>>    Enfin, Kasparov gave the computer 9 stones ahead and yes the computer
won.
>>    How surprisingly. At the internet chessprograms already win that kind
>>    of positions since many years.
>
>a) I don't think DB ist "9 stones" weaker than Kasparov.
>   Even if DB wins only 2/10 games it's just two stones weaker!

the freeware Crafty program is without doubt not a weak program. 
It's blitz elorating is around 2900-3000, which objectively means 
it would win from the strongest GM's on this planet at blitz. In fact it
already has did this.

Note that kasparov's rating which is the highest near humans is around
2800.

This can of course hardly be compared to blitz strength, because that 2800
of kasparov is based on tournament games, and that 2900-3000 is based on less
serious games (humans cannot make money by beating computers; if you want
to know how strong a pro is, play him for money!).

Anyway, in go giving up 9 stones means giving up 9 moves.
In chess giving up moves is also possible by doing some horrible moves.

Now it's good to realize that in chess the development of all things is
way beyond computer go. Way more money and effort has been invested in it,
which means that i have for example a huge book containing 7.7 million moves
based on hundreds of thousands games.

However this book i read in very recently and i didn't remove all crap 
automatically nor by hand, that's for the next months.

This means that some of the games read in are really horrible and causing
the program to play horrible lines sometimes.

Now in chess giving up 9 moves to your opponent from the start of the game
isn't possible. that means that you are mated when you give up the 4th or 
5th move, depending on what side you play.

Giving up 1 move in chess is sometimes possible. Giving up 2 moves is in the 
short term losing. especially if you also move the pieces to a worse place,
which happens.

Here a chess game from my own program against crafty.
Diep (account name DoctorWho run by Lonnie Cook) ran at 
a P200+ computer, crafty at a 4 processor pentium pro machine
with 512 MB sdram:

[Event "ICC 05/24/1998"]
[Site "Internet Chess Club: chess.lm.com"]
[Date "1998.05.24"]
[White "DoctorWho"]
[Black "Crafty"]
[Result "1-0"]
 1. h3 e5        2. a3 d5        3. c4 c6        4. cxd5 cxd5    5. d3 Bd6    
 6. g4 Ne7       7. Nc3 Nbc6     8. Bg2 O-O      9. Bxd5 Nxd5   10. Nxd5 Bxa3 
11. Rxa3 Qxd5   12. Nf3 Qd6     13. Be3 Be6     14. O-O h5      15. Qa4 hxg4  
16. hxg4 Bd5    17. Nh4 Qe6     18. f3 Rfc8     19. Kf2 Rd8     20. Rh1 Rd7   
21. Nf5 Rad8    22. b4 b6       23. b5 Na5      24. Bd2 Nb3     25. Bc3 Nc5   
26. Qc2 f6      27. d4 exd4     28. Nxd4 Qd6    29. Qh7+ Kf8    30. Bb2 Re8   
31. Qg6 Bg8     32. e3 Qd5      33. Rh8 Re5     34. Qh7 Qf7     35. e4 Re8    
36. Bc3 Nd3+    37. Kg3 Qc4     38. Nf5 Nc5     39. Nh6 Nxe4+   40. fxe4 Ke7  
41. Qxg7+ Kd8   42. Bxf6+ Kc8   43. Qg6 Qxe4    44. Qxe8+ Qxe8  45. Rxg8 Qxg8 
46. Nxg8 Rd1    47. Rxa7 Kb8    48. Rh7 Rg1+    
{Black resigns} 1-0

So giving up many many stones is easy.

>b) today's computer-go programs wouldn't even win against "Kasparov"
>   if he gave them 18 stones ...

That's very true. Because the branching factor in chess is so small and
because centralization in chess is in 90% of the games so very important
these 2 facts combined together give chess programs at certain terrains 
already a big edge.

The second reason is of course the huge effort put in computerchess.

Most questions asked here and a lot of discussions in this computer-go group
are kind of naive from my viewpoint, as these discussions have been done 
years ago in computer chess.

Same for implementing several things like knowledge and combining
local searches and global searches. 

On the other hand, my interest in computer go and computer draughts have
lead to several decisions i took about how to write my program which i still
think are in the long term giving my chessprogram a big edge, 
where lots of other chessprograms are becoming
hopelessly outdated. One of them even won the last 
years micro world title.

For example, if i allow my program to search for a night, then it gets
selectively a search depth which can be compared to a brute force search
depth of about 18-20 ply, extending some local tactical variations up to
56 ply (which is my stackdepth otherwise it would be more). 

Deep blue gets by the way just 11 ply with another 4 ply added, so 15 ply 
main search depth.

So with a smart selective way of searching i see the same and even deeper,
than the on old methods based Deep Blue, not to mention how outdated
even older programs already are.

Now suppose developments would lead to smart way of searching which for
go-programs means that they're getting 18-20 ply and some tactical 
variations up to 100 ply, at analysis level?

In certain terrains i hear already that go programs with few ply main search
depth already play reasonably in some positions, just like chess programs do.

In other positions they simply play both horrible. Right now the % of
positions
where go programs play horrible is definitely way way bigger than chess
programs. 

Would huge main search depth lead to closing that gap in the same way?

So not the fact how doable is my question, but my question is: how would
that 'deep believe go program', shortly DB-go program play do you think?

And even more interesting, where do you base your opinion on?

Vincent

>Christoph

Vincent Diepeveen