The program learns to play on videos
The program plays simple tic-tac-toe games, four in a row (Connect 4) and Gomoku - and wins over a person. It would seem, nothing interesting, if it were not for one important nuance - this program does not know the rules! More precisely, she learned them from scratch by watching two-minute videos of people playing games.
Lukas Kaiser (Łukasz Kaiser) from the University of Paris Diderot has written a C ++ program that breaks down the video frame by frame, removes all unnecessary (hands of people) from them - and gets a list of consecutive positions.
Based on these data, the algorithm (on OCaml) makes up the base of allowed moves and the list of winning / losing / unresolved positions, then generates a set of logical formulas like ∃x1Q (x1) x0 (C (x1, x0) ∧ x0 = e1). Both modules are integrated into the free program for playing board games Toss .
On a regular laptop, it takes 28-74 seconds to process a two-minute video, depending on the game. Then another couple of minutes for drawing up a set of rules - and she is ready to play. True, to learn how to win, you need to watch not one video, but more. During the testing program, she processed 25 tic-tac-toe videos, 25 - four in a row and 17 - in gomoku. If the game is more simple, then less is enough.
')
It would seem nothing complicated, but just imagine that such systems will become more complex and sophisticated - and will be able to learn more complex games and strategies, just by watching the actions of humaninsects .
Actually, Lukash Kaiser himself writes that he chose board games as the object of application of his program, “because here the number of visual objects is relatively small, while there is a sufficient variety of actions. In addition, games are a natural model of many relationship scenarios in the real world, which makes the result meaningful in a wide context. ”
For more information about the algorithm, see the scientific work: Learning Games from Videos Guided by Descriptive Complexity (PDF), or on the video recording of his presentation from the Third Conference on Artificial General Intelligence.
Lukas Kaiser (Łukasz Kaiser) from the University of Paris Diderot has written a C ++ program that breaks down the video frame by frame, removes all unnecessary (hands of people) from them - and gets a list of consecutive positions.
Based on these data, the algorithm (on OCaml) makes up the base of allowed moves and the list of winning / losing / unresolved positions, then generates a set of logical formulas like ∃x1Q (x1) x0 (C (x1, x0) ∧ x0 = e1). Both modules are integrated into the free program for playing board games Toss .
On a regular laptop, it takes 28-74 seconds to process a two-minute video, depending on the game. Then another couple of minutes for drawing up a set of rules - and she is ready to play. True, to learn how to win, you need to watch not one video, but more. During the testing program, she processed 25 tic-tac-toe videos, 25 - four in a row and 17 - in gomoku. If the game is more simple, then less is enough.
')
It would seem nothing complicated, but just imagine that such systems will become more complex and sophisticated - and will be able to learn more complex games and strategies, just by watching the actions of human
Actually, Lukash Kaiser himself writes that he chose board games as the object of application of his program, “because here the number of visual objects is relatively small, while there is a sufficient variety of actions. In addition, games are a natural model of many relationship scenarios in the real world, which makes the result meaningful in a wide context. ”
For more information about the algorithm, see the scientific work: Learning Games from Videos Guided by Descriptive Complexity (PDF), or on the video recording of his presentation from the Third Conference on Artificial General Intelligence.
Source: https://habr.com/ru/post/147916/
All Articles