Re: returning results for partial downloads

Serguei Osokine wrote:
> 	Perhaps I was not clear enough. Let's say there are only two
> types of clients. Utility maximizers were never invented. In this
> case, it looks like your success metrics does not work, benefiting
> leechers for no good reason. This makes me doubt the utility of such
> metrics as a whole - if it does not work in a simple model situation,
> how do I know that it will behave itslef correctly in a complex
> scenario? Shouldn't there be a measure of algorithm success that
> would work in all known field of situations?

Sorry, I see what you mean now. There are two goals: the algorithm must
be beneficial to users who adopt it, otherwise it won't be adopted, and
it must be beneficial to the network as a whole. So there are two
metrics: relative download speed obtained by users of the algorithm
compared to users of other algorithms, and total download speed when
everyone uses the algorithm.

I did some experiments a while back with mixtures of three strategies:
free riders (never upload), altruists (upload to all peers equally) and
reciprocators (maximise utility). The results are in section 6 of this
tech report if you're interested:

http://www.cs.ucl.ac.uk/staff/mrogers/transfer-report.pdf

I've also run some evolutionary simulations that show similar results:
from almost any initial conditions, the population stabilises to a
randomly fluctuating mixture of reciprocators and altruists, with free
riders almost eliminated. But I haven't looked at how utility
maximisation competes with other selfish strategies like stochastic tit
for tat.

Cheers,
Michael