I’ll look into an important phenomenon on the chess landscape, AlphaZero, who just released more match games played against the powerhouse publicly available program, Stockfish.
The games were from a private match of 1,000 games that resulted in 155 wins for Alpha Zero, 6 for Stockfish, and 839 draws. The parameters of the match were somewhat arbitrary; AlphaZero played a somewhat older Stockfish 8 version, the allotted time for the competitors was chosen by AlphaZero’s operators, and it’s not entirely clear when the games were played. In any case, I’m more interested in what kind of chess we can see in these games then issues of relative strength. I have only plied through a handful of the games released, and they certainly show a mixed bag. Garry Kasparov has praised AlphaZero’s ”style,” namely the way it plays aggressively rather than maneuver around tediously. Others have chimed in positively as well, Matthew Sadler suggesting that we can all learn from AlphaZero’s strategic ideas. But there is certainly a great degree of mind-numbing maneuvers in several of the games. A whole lot of repetitions bump up the length of almost every game. Yet when AlphaZero found opportunities to attack, it did so with what humans would consider great gusto and great risk.
In my new book Better Thinking, Better Chess, I wrote a chapter on the relationship between material and initiative. I pointed out that grandmasters consider material, especially the pawn count, to be just one feature of chess and will often sacrifice or the initiative, whether or not they see an immediate payoff. Well, the first game we will look at shows this philosophy but turned up several notches!
Stockfish was not exactly defenseless, scoring something in the vast majority of the games. In the next game, the computers go toe to toe and come out even, though along the way we see a fascinating sacrifice followed by a typical computer elongation of the game.
AlphaZero is said to exhibit a human-like intuition in its games. That’s because we are not able to find our way to the justification of its decisions, in the way traditional alpha-beta engines provide us with principal variations to confirm its moves. I’ll have to conduct deeper study after I download the games to get a better feel, but I’m still not sure how much in these games we can understand enough to apply to our own games.