What is Watson?

Discussion of anything and everything relating to chess playing software and machines.

Moderators: bob, hgm, Harvey Williamson

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
Terry McCracken
Posts: 15844
Joined: Wed Aug 01, 2007 2:16 am
Location: Canada

What is Watson?

Post by Terry McCracken » Thu Feb 17, 2011 9:55 pm

Terry McCracken

kgburcham
Posts: 2016
Joined: Sun Feb 17, 2008 3:19 pm

This is Watson?

Post by kgburcham » Thu Feb 17, 2011 10:03 pm


User avatar
Leto
Posts: 2028
Joined: Thu May 04, 2006 1:40 am
Location: Dune

Re: This is Watson?

Post by Leto » Thu Feb 17, 2011 11:48 pm


Frank Quisinsky
Posts: 4852
Joined: Wed Nov 18, 2009 6:16 pm
Location: Trier, Germany
Contact:

Re: What is Watson?

Post by Frank Quisinsky » Thu Feb 17, 2011 11:57 pm

The programming partner from Microsoft and Sherlock Holmes I think, or not?

Cloned in much chess programs, because the error messages are often the same :-)

Best
Frank
I like computer chess!

NATIONAL12
Posts: 305
Joined: Sat Jan 02, 2010 10:31 pm
Location: bristol,uk

Re: What is Watson?

Post by NATIONAL12 » Fri Feb 18, 2011 12:04 am

wonder what brought this subject up. :)

Vinvin
Posts: 4378
Joined: Thu Mar 09, 2006 8:40 am
Full name: Vincent Lejeune

Re: This is Watson?

Post by Vinvin » Fri Feb 18, 2011 12:11 am

Many thanks, Leto !!

jplchess
Posts: 115
Joined: Mon Jan 25, 2010 6:13 am

Re: This is Watson?

Post by jplchess » Fri Feb 18, 2011 12:21 am

To the computer chess community:

http://www.youtube.com/watch?v=d_yXV22O6n4&NR=1

This computer learns and knows puns.

CCC has someone talking about IBM Watson almost 2 years ago.

In more related news, does computer chess software learn from a mistake that has a relatively decent time control and number of plies?

Jonathan Lee

Dann Corbit
Posts: 10124
Joined: Wed Mar 08, 2006 7:57 pm
Location: Redmond, WA USA
Contact:

Re: This is Watson?

Post by Dann Corbit » Fri Feb 18, 2011 1:25 am

jplchess wrote:To the computer chess community:

http://www.youtube.com/watch?v=d_yXV22O6n4&NR=1

This computer learns and knows puns.

CCC has someone talking about IBM Watson almost 2 years ago.

In more related news, does computer chess software learn from a mistake that has a relatively decent time control and number of plies?

Jonathan Lee
Some programs have position learning.
Some programs have evaluation learning.
Some programs have book learning.

Terry McCracken
Posts: 15844
Joined: Wed Aug 01, 2007 2:16 am
Location: Canada

Re: This is Watson?

Post by Terry McCracken » Fri Feb 18, 2011 1:50 am

Dann Corbit wrote:
jplchess wrote:To the computer chess community:

http://www.youtube.com/watch?v=d_yXV22O6n4&NR=1

This computer learns and knows puns.

CCC has someone talking about IBM Watson almost 2 years ago.

In more related news, does computer chess software learn from a mistake that has a relatively decent time control and number of plies?

Jonathan Lee
Some programs have position learning.
Some programs have evaluation learning.
Some programs have book learning.
Do any have all three and are they any good?
Terry McCracken

Dann Corbit
Posts: 10124
Joined: Wed Mar 08, 2006 7:57 pm
Location: Redmond, WA USA
Contact:

Re: This is Watson?

Post by Dann Corbit » Fri Feb 18, 2011 2:31 am

Terry McCracken wrote:
Dann Corbit wrote:
jplchess wrote:To the computer chess community:

http://www.youtube.com/watch?v=d_yXV22O6n4&NR=1

This computer learns and knows puns.

CCC has someone talking about IBM Watson almost 2 years ago.

In more related news, does computer chess software learn from a mistake that has a relatively decent time control and number of plies?

Jonathan Lee
Some programs have position learning.
Some programs have evaluation learning.
Some programs have book learning.
Do any have all three and are they any good?
Some have all three.

Evaluation learning via td-lamda and td-leaf are the least effective, from what I have seen (custom evaluation seems to do better).

Position learning is 100% effective if you hit the same position again (but this has a surprisingly low probability). Suppose (for instance) that you are playing a chess game and encounter this position:
[d]2r3k1/4ppb1/2P5/4P2p/2R3p1/1p6/1B4PP/5K2 w - -
Your chess engine makes a bad move and writes out a record that stores the correct value after the opponent's move. The odds that you are going to play this move again are basically zero, unless it is near the origin (in which case it is probably a book move).

Book learning is effective, but it gets poisoned if you run lots of blitz games. IOW, if you run nothing but correspondence time control with an ultra strong program, eventually you will get a nearly perfect book. However, it will take a very long time to get there.

The problem with blitz book learning is that a computer program can easily make a mistake because of a shallow search. The book gets updated with this new wrong answer. So the engine will avoid moves that have this move now flagged as bad, but potentially it is a good move.

My synopsis:
The only one of these strategies that actually work well in practice is book learning and the only way that book learning will work optimally is if you run at very, very slow time control and with a very strong engine.

Eventually, computers will become fast enough that blitz time control will safely update book learning, but I guess that this is still some years away.

I think that there is a long way to go before computers utilize chess statistics properly in learning and move selection. It is actually something that I am actively working on.

Post Reply