History of Segmentation/Prioritizing

Robo Home | Changes | Preferences | AllPages

Difference (from prior major revision) (author diff)

Changed: 76c76

Essentially it's taking the entropy of the peaks (how much the peaks vary across the segment), with artificially reduced entropy where hit rates are low. I'm not sure how to explain exactly how I came up with this model, however I think it's an advantageous one, because it has the following properties:

Essentially it's taking the entropy of the peaks (how much the peaks vary across the segment), with artificially reduced entropy where hit rates are low. I think it's an advantageous model because it has the following properties:

Changed: 79c79

So what to other people think of this model here? -- Rednaxela

The reasoning behind creating the model this way, as that principals on entropy and information gain as described above, are only meaningful if all information is of (at least approximately) equal value. As an alternative to "information gain" this method goes for what I call "useful information gain" which multiplies by a "usefulness" value. So what information is useful? In the context of Robocode, useful information is information that leads to (or dodges) hits, and thus it makes sense to use the hit rate as the multiplier on how much each peak adds to entropy. So what to other people think of this model here? -- Rednaxela

	Revision 12 . . March 20, 2008 0:40 EST by Rednaxela [Elaborate on the reasoning behind this hybrid idea]
	Revision 11 . . March 19, 2008 17:31 EST by Rednaxela [Idea for an accurate entropy/hit% hybrid model]
	Revision 10 . . February 4, 2008 23:03 EST by Voidious [comment]