Google's self-learning AI AlphaZero masters chess in 4 hours
Video Statistics and Information
Channel: ChessNetwork
Views: 1,461,833
Rating: 4.8631191 out of 5
Keywords: alphazero, alpha zero, alpha zero vs stockfish, alpha zero chess, alphazero vs stockfish, alpha zero vs stockfish 8, chess, chessnetwork, chess network, chess ai, google's deepmind, deep neural networks, general reinforcement learning algorithm, best chess player ever, alphazero masters chess, google's alphazero masters chess, google's deepmind ai
Id: 0g9SlVdv1PY
Channel Id: undefined
Length: 18min 10sec (1090 seconds)
Published: Thu Dec 07 2017
Please note that this website is currently a work in progress! Lots of interesting data and statistics to come.
Interesting to see that in some games it won with ambitious and aggressive sacrifices, and in some games like this it just wore stockfish down positionally until it ended up in a winning position. I know it's basically a meme at this point but this thing really does play like a human.
I've never really liked looking at engine games but these games have seem pretty instructional to me. I liked Jerry's analysis here, especially noting about having the knight block the pawn instead of the bishop.
I can't wait to see the Starcraft iteration.
It almost looks like it doesnβt have a style at all. Itβs like - chess evolution has always been about some people showing that the current understanding of chess is insufficient - People grew out of romantic era of throwing all out attacks from move 1, capablanca introduced harmony and positional maneuvering, Tal showed that level of defense of the time was too low and so on.
For me it seems that AlphaZero doesnβt discriminate any aspect of chess. For it, material, piece activity and pawn structures etc are just another aspects of positional chess that sometimes are important and sometimes not. Itβs a dynamic game after all. Reviewing the white games reminded me of Jacob Aagaard quote from one of his books that goes something like: "To be a strong positional player, it is required to have somewhat liberal approach to material. Material is just another factor of positional chess. What good is a material lead of a Bishop and a pawn versus a lone king, if itβs the classic wrong corner situation, and the defending king gets to the corner? Clearly the king position compensates for the material disadvantage.β
Of course no human can ever have the tactical vision to prevent stockfish from untangling itself from the positions it got in the French/queenβs indian games, but clearly AlphaZero assessed the piece activity and space advantage to be more important than material, and it certainly seems like it was right. From what I know, it might not need point counts for pieces or any other defined factor for assessing positions, it maybe simply assesses the positions as they are as a whole and goes for the most promising variation. Material, space advantage and piece activity are, after all, human concepts that help in making sense of chess, the AI might not need them at all.
I love Jerry but I found this analysis kinda lacking. There are some good points, like the rerouting of the bishop/knight, but for the most part it felt like he was just telling us the moves that were played.
Maybe I'm the stupid one for expecting a deep analysis when the video is only like 15 minutes long, but it's totally unclear to me which moves by Stockfish were slightly inaccurate and where it went wrong. Then again, the Danny Rensch analysis of game 3 is around the same length but much more thorough.
What kind of timing was used in these games? Standard 2 hours for 40 moves?
Stockfish has deficiencies that no human player can find. But they had to be there.
I want to see Alpha zero try it's hand with law. What would its policies be?
I'm off on a wild tangent but I wonder if they'll eventually use deepmind to decipher the language for other species. Will we eventually be able to communicate efficiently with whales, dolphins, elephants?