Denser sampling only seems useful if the problem domain is in some way smooth - interpolatable. If you run it on a fractal problem domain you just learn more special cases. Chess is fractal.