On Wed, Jul 13, 2005 at 06:07:55PM +1000, Damian Conway wrote:

> The use of summed lineal distance (L[1]) rather than RMS distance (L[2]) 
> probably *isn't* superior as a closeness measure. But it's computationally 
> much simpler (and hence likely to be more efficient), it doesn't suffer 
> from precision issues in "photo finishes", and it is almost certainly 
> easier for the average programmer to predict correctly.

I dispute "much", specifically because the sum of the squares has the same
ordering as the square root of the sum of the squares.

(So we're talking only about 1 extra multiply per class under consideration,
and a possible overflow or need for floating point, rather than the more
direct assumption that ordering by RMS would need a square root evaluated)

But I have no opinion on the underlying merits of L[1] versus L[2], because
I have no experience of MMD.

> That said, I'd have no *particular* objection to an MMD implementation that 
> used RMS inheritance distance as its metric, provided the dispatch 
> performance was not appreciably worse.

Maybe I should cut that paragraph, to avoid re-opening a debate.

Nicholas Clark

Reply via email to