Move Over, AlphaZero: There's a new girl in town
By now, you've probably heard of AlphaZero, Google's incredible chess project. Last year, it took the chess world by storm. Using 'artificial intelligence' - at least that's what the news articles claimed - AlphaZero taught itself chess in less than a day and then demolished the world's strongest chess engine in a 100-game match.
There are several small ambiguities in this summary. First, Google's hardware was enormous, so it may not have been a fair fight (could you beat Louis Hamilton if he was driving a go-cart?). And second, it's not really 'AI' technology being utilised, but rather a neural-network approach to computations. Don't ask me how it works, because I don't know. But I understand enough to know that utilising Monte-Carlo simulations is quite a different thing from Scarlett Johansson in Her.
The engines match happened to occur during the London Chess Classic last year, so we got to hear the first reactions of the world's elite. They were impressed. Some of them were mighty desperate to get their hands on the technology (perhaps without realising that it involved a computer setup the size of a small house). Alas, it seemed Google's megamonster was out of reach for the average chess pundit.
Until now. Recently, an ambitious project called LeelaChessZero was started. It involves the same general idea of AlphaZero, in that the computer plays millions and millions of games against itself in order to learn (without any direct human guidance). But in order to compensate for the computing power required of Google, this project has reached out to the wider (and surprisingly altruistic) chess community. Right now, as you read this, thousands of games are being played by LC0 on computers and servers around the world. Each training game contributes to the wider neural network, as if the program is being fed.
And in fact, you don't even need to slow down your own computer to be a part of this. You simply need an internet connection in order to help out, by using - and I love this irony - Google's own online 'Colaboratory'. It's completely free and takes 2 minutes to set up. Go to this helpful guide to find out how. How strong is LC0 now? It's closing in on the top chess engines in the world, though it still has a fair way to go. On reasonably hardware, it's currently performing at around 3000 ELO (the world's best is estimated around 3500). Most interestingly, it doesn't play like a normal engine: LC0's tactics actually took a long time to get going, while its positional play grew rapidly. Ironically, it was almost learning like a human: The general strategies were there in the early days, but it would still fall for mate-in-one!
The other really interesting (and somewhat human) thing is that it hasn't been smooth sailing all the way. In the past month, LC0's strength actually DECREASED from playing more training games. Completely puzzling to me, though there's a lot of technical discussion about this on the forums that I don't begin to understand. Leela's adolescent crisis
I don't know where this project will go in the future, but it's quite exciting. Meanwhile, I've got Google Colab running in the background, just so I can say I was a part of it.