What is Watson?

[Terry McCracken]

I can't believe Jeopardy! didn't have a category about computers with A.I. (fictional or non) with answers from movies like War Games, 2001 A Space Odyssey, Terminator, Star Trek, etc.

I can....

[gaard]

http://www.theatlantic.com/technology/a ... son/71584/

Kasparov on IBM's new "toy".


http://blog.reddit.com/2011/02/ibm-wats ... -your.html

The research team finally answers reddit's questions.

[rbarreira]

(...) chess computers. They play fantastically well in maybe 90% of positions, but there is a selection of positions they do not understand at all. Worse, by definition they do not understand what they do not understand and so cannot avoid them.

It seems Kasparov still hasn't given up the idea that humans are superior to computers at chess.

[Leto]

(...) chess computers. They play fantastically well in maybe 90% of positions, but there is a selection of positions they do not understand at all. Worse, by definition they do not understand what they do not understand and so cannot avoid them.

It seems Kasparov still hasn't given up the idea that humans are superior to computers at chess.

The statement you quoted doesn't support that claim, it simply says chess computers don't do well in a selection of positions. Humans probably play fantastically in 10% of positions.

[rbarreira]

(...) chess computers. They play fantastically well in maybe 90% of positions, but there is a selection of positions they do not understand at all. Worse, by definition they do not understand what they do not understand and so cannot avoid them.

It seems Kasparov still hasn't given up the idea that humans are superior to computers at chess.

The statement you quoted doesn't support that claim, it simply says chess computers don't do well in a selection of positions. Humans probably play fantastically in 10% of positions.

But he also says that computers can't avoid those positions they can't play in. Which immediately implies that humans could win by tricking computers into those unavoidable positions they don't play well, according to Kasparov's assumptions.

[Dirt]

This was not a fair game whatsoever .

Yeah, the computer has a faster reaction time. How are you going to fix that?

In any case, just getting to the point that the reaction times mattered was quite an accomplishment. Congratulations to IBM.

[JManion]

This was not a fair game whatsoever .

Yeah, the computer has a faster reaction time. How are you going to fix that?

In any case, just getting to the point that the reaction times mattered was quite an accomplishment. Congratulations to IBM.

Unplug the computer, that usually kills its reaction time.

[Terry McCracken]

This was not a fair game whatsoever .

Yeah, the computer has a faster reaction time. How are you going to fix that?

In any case, just getting to the point that the reaction times mattered was quite an accomplishment. Congratulations to IBM.

Unplug the computer, that usually kills its reaction time.

In 40 years you may find that old trick rather hard to pull...

We now know can build a machine that can play chess and carry a coherent converstion.

HAL in 2051? Who knows?

[rodolfoleoni]

....................

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).

......................

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.

I don't fully agree about low probability of a learned position to be hit in search. Let us suppose that the learning engine runs a special book where it always plays the same opening lines... E.g. always english as white, always Caro-kann, Nimzoindian as black. Let us suppose there's only one move for learning engine in that book, while it includes all possible opponent reply. That book would consist of few hundreds variations. The engine would learn the related middlegame position fast enough.

I built a book like that some years ago, and it showed to be good for position learning. The limit of that book was that I'm not a too strong player (about 2000 ELO) so it's not a high quality book, but I guess if a master rated player would do a similar work it could help learning much better.

There's a new feature of the Baron 3.X which shows to be effective with learning. It's called "deeplearn" and it consists of building a tree of learned positions from a root position (which can be the position the engine goes out of book). I had no much time to test it, but it seems great. Richard (Pijl) told me he has some new interesting ideas about it for Baron 4.0 (work in progress).

[rodolfoleoni]

....................

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).

......................

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.

I don't fully agree about low probability of a learned position to be hit in search. Let us suppose that the learning engine runs a special book where it always plays the same opening lines... E.g. always english as white, always Caro-kann, Nimzoindian as black. Let us suppose there's only one move for learning engine in that book, while it includes all possible opponent reply. That book would consist of few hundreds variations. The engine would learn the related middlegame position fast enough.

I built a book like that some years ago, and it showed to be good for position learning. The limit of that book was that I'm not a too strong player (about 2000 ELO) so it's not a high quality book, but I guess if a master rated player would do a similar work it could help learning much better.

There's a new feature of the Baron 3.X which shows to be effective with learning. It's called "deeplearn" and it consists of building a tree of learned positions from a root position (which can be the position the engine goes out of book). I had no much time to test it, but it seems great. Richard (Pijl) told me he has some new interesting ideas about it for Baron 4.0 (work in progress).