4 ms·
So you've now moved the computational complexity to the fitness function, which needs to consider future moves as well?
by Kjeldahl 11y ago
So you've now moved the computational complexity to the fitness function, which needs to consider future moves as well?
- oh_sigh 11y agoNot necessarily. If there are factors of a board position that make it strong regardless of future moves, then why can't that information be encoded in the fitness function without having to play future moves forward?
- Kjeldahl 11y agoSure, but how would you know? Sacrificing officers in chess is a typical example. Negative expected value short term, but typically long term positive when done intentionally. But generally, if you can come up with a robust fitness function, genetic algorithm is one method, although typically there are better alternatives giving faster results. In my experience, genetic algorithms can be a great start to get something but typically slow. Which matters if you need to simulate a few billion games... Interesting article anyway, thanks.
- togelius 11y agoThe fitness function is pretty simple, essentially a weighted piece counter. It's true that it does not look ahead and take future states into account so might fall into a trap, but empirically it's doing very well!