> From: MULLER Guillaume <[EMAIL PROTECTED]> > Date: Fri, 10 Dec 2004 09:33:39 +0100 > To: [EMAIL PROTECTED] > > Hi all, > > Right, I would have cited Dellarocas' papers also because he is the only=20 > one I know that worked on this subject. > > However, IMHO, his claim that size of history doesn't matter is false.=20 > He took this conclusion in very a specific domain that is eBay-like=20 > market-places with very specific assumption (cf. cited paper). > > My idea is that size of history DOES matter. Let's imagine a system=20 > (even eBay-like) where every agent *knows* that the history is a list of=20 > the X last encounters experiences. Then it is easy to see that cheating=20 > 1/X times is a strategy that pays off (particularly in systems where=20 > ratings might be noisy). > > IMHO, the key point with respect to the history is that others should=20 > not be able guess its size. If it has a fixed size, I believe it doesn't=20 > matter if (and only if) other can guess its size (and therefore cannot=20 > use strategy as described above). > > However, I'm sorry I didn't have time to make any experimentations, but=20 > I'd like to hear if anybody has.
(1) You'll never eliminate cheating. (2) Making the size of the history file a secret is probably unworkable. Better to make deletion from the history non-deterministic, so the longer a record has been been in the list the more likely it is to get dropped. A potential cheater would never be certain when the incriminating evidence would be gone. If which records were disreputable was known then their lifetime could be extended. cheers, Tim