字幕表 動画を再生する 英語字幕をプリント If you watched my last video, you would know that the game of Go is based on such a simple set of rules, much less complicated than Chess. But why Go computers couldn't beat the best professional Go players until recently, while Chess computers have done it for nearly two decades. One can simply answer that they have fundamentally different approaches of how they think when playing. So, how a chess computer actually think? For a human being, the game of chess involves a great deal of high-level abstract reasoning, pattern recognition, conscious thought and sometimes even psychology. By accumulating experiences games after games, people develop and improve their skills and strategies. Some might argue that we should program the computers to imitate human thinking. In the past, many early researchers of artificial intelligence also agreed with that, and made the mistake we now call the AI fallacy. They assumed that the only way to perform non-routine tasks of invention, creativity and judgment to the standard of human experts is to replicate the approach of human specialists. Chess computers do none of that. In 1997, Deep Blue, the supercomputer of IBM, beat Garry Kasparov, who then became world chess champion in a 6-game series. The machine didn't have intuition or experience like Kasparov, it won by abusing its sheer processing power and massive data storage capability. To simply put, instead of thinking, Deep Blue calculated its way to victory, which is also known as the brute force approach. In this method, a chess engine builds a search tree consisting of every possible future move from both players. In this example tree, white circles represent the chess engine's moves, while black squares represent the opponent's moves. A ply is just one move by one player. The number of children at a position is called the branching factor. Once the tree is created, the chess computer use Minimax algorithm to decide which move is the best, based on an assumption that the opponent has similar thoughts, and will try its best to win. This algorithm was first developed by John von Neumann in 1928, and adapted for chess by Claude E. Shannon in 1950. Back to the example tree, each final position has been assigned a number based on the computer's result, which are 1, 0 and -1 for win, draw and loss. The chess engine works upward from the bottom ply, it chooses the best positions from each possible position black will take, or the maximum. At the next level, it assumes that black will choose the worst possible position for white, or the minimum. In the end, the computer takes the maximum of the top three positions to make the move. After black makes its move, the computer creates a new tree and goes through this process again, rinse and repeat till the game ends. Programmers immediately realized a huge problem of Minimax. It's impossible to search the entire tree of a chess game, where the average branching factor is approximately 35 and the average depth of plies is about 80. That means the estimated size of search tree for chess would be 10^123. If this number sounds incomprehensible for you, just remember there are only about 10^75 atoms in the universe. This problem is commonly solved by using a fixed-depth search, or limiting how far ahead the chess engine will look. It will then analyze each position in the last ply to decide whether that position is good or bad by using a static evaluation function. This is one of many deciding factors distinguishing between good and bad chess engines, as the good ones usually have a much more accurate and complicated function. For example, a mediocre chess engine might just compare the number of pieces each side has, while the better one might apply a weight to each type of piece, because obviously a queen is much more valuable than a pawn. However, no matter how complicated the function gets, each position is eventually assigned a number representing how good that position is. Programmers again realized another serious problem, the horizon effect, first described by Hans Berliner. Imagine the chess engine only searches to a depth of 8-ply. At ply 8, the computer's Queen captures an enemy's pawn. It sounds great until you realize that pawn is defended by another pawn, so the opponent will take the computer's Queen the next move at ply 9, but the chess engine can't see that, because it stopped searching at ply 8. Programmers can't just simply increase the depth limit to avoid this, because there will be always a horizon, no matter how deep it is. An algorithm called "Quiescence Search" was invented to tackle this problem. After reaching the final ply, the chess engine enters a special search mode that only expands certain types of moves, and only finishes when it gets to a quiet and relatively stable position. There is still another problem. The fixed-depth search reduces the strain on processing power, but it also compromises the quality of decisions due to lack of depth. To be able to increase search depth and look further, programmers use the Alpha-Beta pruning algorithm to lower the branching factor. Basically it allows the computer to discard many tree branches because the computer can see early on that the branch it's going to check is worse than another available checked branch. A good Alpha-Beta pruning can significantly reduce the branching factor from 30 to only 5. The fixed-depth Minimax, with Alpha-Beta pruning and Quiescence search have been the cornerstones of almost all chess engines, from the famous 1997 Deep Blue to the modern ones as Stockfish, Komodo and Houdini. Nowadays, thanks to new algorithmic developments, combined with advanced hardware, Stockfish, the strongest chess engine, only needs an iPhone instead of a tower full of CPUs, to beat any chess grandmaster. That said, this doesn't mean computers are becoming intelligent like humans. Their dominance in chess and other board games only prove that, machines became fast enough to exceed the playing capacity of the best human player. Go computers might be an exception, but it's a different story, which I hope to talk about in another video. As a creator of Deep Blue has said, "Assume that we were playing Kasparov at the World Trade Center, and 9/11 hit. Kasparov would've run like hell. We would've run like hell. Deep Blue would just sit there... computing."
B1 中級 米 チェスのコンピュータはどう考えているのか? (How Do Chess Computers Think?) 86 6 Vinh Nguyen に公開 2021 年 01 月 14 日 シェア シェア 保存 報告 動画の中の単語