News for hackers. Stuff that matters.
Re:A way to deal with enormous branching factors in strategy... by Kjeldahl on Friday March 25, 2016 @02:41PM (#11361866)
So you've now moved the computational complexity to the fitness function, which needs to consider future moves as well?
Re:A way to deal with enormous branching factors in strategy... by oh_sigh on Friday March 25, 2016 @03:02PM (#11361976)
Not 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?
Re:A way to deal with enormous branching factors in strategy... by Kjeldahl on Friday March 25, 2016 @03:13PM (#11362038)
Sure, 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.
Re:A way to deal with enormous branching factors in strategy... by togelius on Saturday March 26, 2016 @12:31AM (#11364432)
The 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!