Well there’s been a frisson of excitement in the chess and AI world lately with the extraordinary performance of AlphaZero – essentially the computer that mastered the game Go – a game which proved, despite the relative simplicity of its rules, a much harder nut to crack than chess.
In any event last week a tweaked version of the computer that succeeded in Go was taught the rules of chess but then, without any further instruction, trained itself for four hours by playing games with itself. Then over 100 games, AlphaZero beat one of the best chess engines in the world – Stockfish – winning 25 games as white, 3 as black, and drawing the remaining 72.
Whereas the existing engines employ a mix of techniques with humans training them in various position evaluation techniques and brute computing power being used to search vast move trees, AlphaZero puts its effort into training itself to use better algorithms. Once trained, like a human, the algorithms it’s devised economise on the computing power needed to play.
Though one of the emerging stylised facts of AI is that good collaboration between humans and computers beats computers on their own, at least here the humans set the program up and it then does all the rest. Still it’s still a collaboration and the nice thing is that, this new form of collaboration produces more ‘human like’ chess and more entertaining chess also.
Thus, although it seems odd to say it’s more human given that AlphaZero takes chess even further from the capabilities of humans on their own, it has been the case for some time that, despite their sophistication, chess engines typically play a dull, excessively ‘technical’ kind of game in which long range strategic or ‘position’ considerations are downplayed to raw technique as they gnaw away at some small weakness of their opponents and sometimes grind them down (computers can be trained not to mix metaphors I expect but I’m not a computer. Anyway gnawing is a kind of grinding).
AlphaZero plays like a human in the sense that it’s got much more ‘positional’ savvy. It will give up material for an advantage that looks pretty speculative. One thing that separates a good from a bad player is that a good player can take an initiative and build on it. But if you aren’t technically good (like for instance me!), some inaccuracy in your followthrough will enable your opponent to stabilise the situation and you’re then just down material and you lose. Even amongst very good players, the sacrifice of more than a pawn for long-term position advantage is a rare and fine thing to watch.
Looking at AlphaZero’s games the computer can gain a small edge and, by virtue of its extreme accuracy it doesn’t give it away. In most of the games where I’ve seen it win, it usually gives up a pawn or two and then twenty moves later – even though Stockfish has thought it’s position was cramped but pretty good – the opponent discovers that it’s being completely asphyxiated.
In the game below, one of the best, AlphaZero sacrifices pawns and a piece with no clear pathway to winning. But of course there is a clear pathway to winning. It’s just takes a long time to see it. And judging from what we know of AlphaZero’s programming and it’s much lower exploitation of computer power to calculate specific variations, this monster has positional understanding that passeth all human understanding. As Magnus Carlsen’s second said when looking at these games. I’ve often wondered what it would be like if a superior species came to earth and started playing chess. Now he knows.