Let Gandalf Go, You Monsters
July 16, 2019 6:02 AM Subscribe
Back in the day, you could find atrocious behavior even in serious chess engines. They were very materialistic, and a strong player could trick them with positional sacrifices.
Even in 2000, Kramnik crushed the program Junior with a Stonewall Atttack, which is a kind of amateurish opening from the POV of a player of Kramnik's level. There was a "horizon effect" at work with this sort of opening, and programs often could not see that White was slowly building a fatal attack. I don't think it works anymore, but the best chess engines still have trouble seeing that certain kinds of "fortress" positions are draws.
posted by thelonius at 6:46 AM on July 16, 2019 [1 favorite]
Even in 2000, Kramnik crushed the program Junior with a Stonewall Atttack, which is a kind of amateurish opening from the POV of a player of Kramnik's level. There was a "horizon effect" at work with this sort of opening, and programs often could not see that White was slowly building a fatal attack. I don't think it works anymore, but the best chess engines still have trouble seeing that certain kinds of "fortress" positions are draws.
posted by thelonius at 6:46 AM on July 16, 2019 [1 favorite]
I'm sure any of these could beat me at chess. I'm that bad. I'm not even sure I can beat Microchess.
posted by scruss at 7:05 AM on July 16, 2019
posted by scruss at 7:05 AM on July 16, 2019
Interesting video. I'd have liked to hear more about the technical details of the neural network design he used. Did he use one that has been used before, or did he pick the layer design himself for certain reasons, that sort of thing.
posted by They sucked his brains out! at 9:22 AM on July 16, 2019
posted by They sucked his brains out! at 9:22 AM on July 16, 2019
For those who were curious, "satranç otomatı" = "chess automaton" in Turkish. Also, some excellent troll physics in the bit about polarization & reflection.
posted by Johnny Assay at 10:15 AM on July 16, 2019 [4 favorites]
posted by Johnny Assay at 10:15 AM on July 16, 2019 [4 favorites]
I'm only a third of the way through the video, but this is amazing
I came here to say the same thing at the same timestamp
posted by fleacircus at 12:55 PM on July 16, 2019 [2 favorites]
I came here to say the same thing at the same timestamp
posted by fleacircus at 12:55 PM on July 16, 2019 [2 favorites]
So the way my phone (and yours) avoids crushing us is basically the same way used in TFV to adjust/measure strength: Pick the best move N% of the time; the rest of the time just do something (close to) random...
What's interesting to me is that this leads to uninteresting games. Who wants to play a genius who'll turn into a total idiot 2.7 times per game? You end up just trying to survive until the inevitable epic blunder, and it's nothing like playing a human.
So now that engines have surpassed us to this ridiculous degree, to me the next question is how to teach a computer to play a convincingly human-like game of chess. I know people involved in this kind of thing for video games... I wonder if anybody has tried to run a chess-based Turing test
posted by kleinsteradikaleminderheit at 1:30 PM on July 16, 2019 [2 favorites]
What's interesting to me is that this leads to uninteresting games. Who wants to play a genius who'll turn into a total idiot 2.7 times per game? You end up just trying to survive until the inevitable epic blunder, and it's nothing like playing a human.
So now that engines have surpassed us to this ridiculous degree, to me the next question is how to teach a computer to play a convincingly human-like game of chess. I know people involved in this kind of thing for video games... I wonder if anybody has tried to run a chess-based Turing test
posted by kleinsteradikaleminderheit at 1:30 PM on July 16, 2019 [2 favorites]
Also, some excellent troll physics in the bit about polarization & reflection.
He has a nice dry sense of humor there. I almost spit my coffee out when that happened.
posted by Avelwood at 2:23 PM on July 16, 2019 [1 favorite]
He has a nice dry sense of humor there. I almost spit my coffee out when that happened.
posted by Avelwood at 2:23 PM on July 16, 2019 [1 favorite]
Yeah, he likes to slip in very dry jokes to his videos
posted by DoctorFedora at 2:57 PM on July 16, 2019 [2 favorites]
posted by DoctorFedora at 2:57 PM on July 16, 2019 [2 favorites]
I've evaluated this as the best of the internet.
posted by Ivan Fyodorovich at 4:36 PM on July 16, 2019 [2 favorites]
posted by Ivan Fyodorovich at 4:36 PM on July 16, 2019 [2 favorites]
to me the next question is how to teach a computer to play a convincingly human-like game of chess
Well, which human are we talking about here? Because obviously humans differ wildly in skill levels. I would imagine you could simulate skill levels by snapshotting a NN at different stages in training, but I am not an AI expert so I may be full of it.
posted by axiom at 9:55 PM on July 16, 2019
Well, which human are we talking about here? Because obviously humans differ wildly in skill levels. I would imagine you could simulate skill levels by snapshotting a NN at different stages in training, but I am not an AI expert so I may be full of it.
posted by axiom at 9:55 PM on July 16, 2019
I imagine that online chess creates huge datasets of matches that are stratified into different skill categories. Training a NN on one of these would work pretty well to simulate humans at a given level.
posted by Ned G at 7:03 AM on July 17, 2019 [1 favorite]
posted by Ned G at 7:03 AM on July 17, 2019 [1 favorite]
« Older Know your Roots: Ube, the Purple Filipino Yam | Tōhokudaigaku burūgurasu dōkō-kai Newer »
This thread has been archived and is closed to new comments
posted by 256 at 6:26 AM on July 16, 2019 [4 favorites]