I was musing over Google’s AI that handily beat Houdini 5 or 6, although there was a great many draws. Then I got to thinking about the creativity of humans. Specifically at crafting chess compositions. I wonder if chess artificial intelligence could ever get to the point where it could not only create chess compositions, but do it in a way that makes it feel like it wasn’t crafted by a machine.
In many chess compositions, there is a theme to solving it. Arguable the harder the problem, the harder it is to finesse the theme. But I’d think many of the composers start crafting a composition based on some theme they have floating in their head. Then they construct it around that theme. I wonder if artificial intelligence would understand that in a chess composition, every piece has some sort of role in the composition, even if it’s very minute.
When I say A.I., mean like google’s chess A.I., in which is just given the basic rules of chess and have it play itself over and over. Could you create a generic A.I, in which you give it the basic rules of chess and the basic rules for chess composition, and have it create really nice chess compositions.
I know someone who works for Google, a tech team leader, I think, in their distributed infrastructure area. We have babbled a bit over time about their A.I. work. They have been working a lot on neural net technology as a basis for it and on algorithms that form a framework learning using that basis. They already had sophisticated hardwired neural net processors they designed/developed that were available on their internal distributed network for use. They used those when they did the Go playing A.I. that wound up beating high level professional Go players readily.
But state of the art network speeds are slow compared to high end computer bus speeds, so they created an A.I. computer system with a lot of neural net processing nodes right on the system bus. They also improved their learning algorithms in parallel. They used chess as a test to see how well it could learn without using human games and theories, but just learned by playing itself. It fared amazingly well. The tech staff got to see how quickly it developed opening theory, …, as there was a lot of internal interest.
When they thought it was ready (a surprisingly short period of time), they then had it play a 100 game match against the latest stockfish. The result of that is public knowledge, as it only won or drew.
We got to talk about this soon after that. He had always felt that neural nets were a good way to do this. I never had enough time and energy to go back and learn about them, but always said you have to show me it is better than what the other approaches do. I had been helping one of the chess program teams back in the mid-70s and also knew some of Ken Thompson’s flirting with chess as a computer problem worth the effort. That was when I first got to know this person when he was a kid just starting grade school. I became his step-father just a bit later. Back then I was actually decent at chess, though I never managed to find time to do tournament chess.
Anyway, after he got to enjoy having been right about neural nets (I was suitably impressed, of course), I said that now they need to attack the next hard problem: how an A.I. can turn that knowledge into human understandable form so it can tutor them, and create tests and such that matches the individual needs at the time. Now chess compositions fit right in with that. So you are right on the button with this.