Hacker News Reader

Alphazero mastered chess by already-implemented self-play.

A slight change in the game rules can have a significant effect on the game and strategy, as anyone that delves into chess variants knows. Is the same true for their implementations?

What is the limitation for state of the art AI in adapting in such a way to rule changes?

View on HN

Turns out DeepMind has already been working on exploring variants with Vladimir Kramnik!

https://www.chess.com/article/view/no-castling-chess-kramnik...

Currently they have to train from scratch, but adapting to new tasks is an active area of research for RL broadly.

No-castling chess didn't need to be trained from scratch. Starting with a trained or partly-trained AlphaZero would certainly have worked, since no-castling chess is a strict subset of chess. The only reason not to is if you are DeepMind and training resources are really cheap.

For more esoteric variants like this, it makes sense to start from scratch, though eventually you could imagine an AI that is good at adapting to all perfect-information two-player games, or even others.

It's a very simple change, since the fundamentals of the game is still the same. It should take no more than a day to change the code and re-train the model.

Reply on news.ycombinator.com