Thu, Oct 3
Incentives for Sharing in Peer-to-Peer Networks by Golle, Leyton-Brown and Mironov
Presented by Jason D. Bakos
Jason D. Bakos [firstname.lastname@example.org]
This paper begins by describing peer-to-peer networking as it relates to popular file sharing
applications, such as the now-defunct Napster. The system allows users to join a network and
share files via download and upload. The mechanism that motivates this system is flawed due to
a “free rider” problem that allows users to benefit from the network without contributing to it.
Users who contribute to the system have altruism as their only motivation to contribute. This
paper proposes several modifications to the mechanism that would eliminate the “free rider”
problem and make the system more successful. These mechanisms are analyzed using a utility
function and at Nash Equilibrium. The proposed mechanisms are shown to motivate users to
contribute. Finally, a simulation is performed using reinforcement learning that that verifies the
author’s earlier conclusions and provides additional information about the nature of the
The first class of mechanisms is called “Micro-Payment” mechanisms. This class of
mechanisms relies to the central database server to maintain upload and download counts for
each user. Each user is credited for uploads and debited for downloads with a fixed amount of
money for each transfer. In this system, users can pay or profit from their activities. This
mechanism is shown to motivate users to share file but is also stricken with other problems,
namely the distastefulness of directly mapping files to currency. A modification to this type of
mechanism is proposed that quantizes payments made for files. This mechanism also has
problems, namely the potential for abuse of the system.
The second class of mechanisms is called “Point-Based” mechanisms. In this mechanism, a
level of abstraction is placed between files and currency. Users spend and gain points by
downloading and uploading, respectively. This mechanism also motivates users to contribute,
but without the disadvantages of the micro-payment mechanism. A modification to this system
is proposed that users may find more appealing. This modification rewards users for sharing
files, rather than the actual uploading of files. This mechanism yields similar results than the
mechanism without the modifications but may be more palatable to the users.
The simulation is run using reinforcement learning, which is suitable for finding equilibrium.
When the learner converges, equilibrium has been reached. The first result was that the
simulation results did not vary when the size of the agents’ action space was changed. Secondly,
the number of shared files the agents shared was seen to change drastically when the utility value
of money was varied. Lastly, when the simulation was run with a mixture of altruistic agents
and non-altruistic agents, it was seen that the non-altruistic agents download and share more
1.The author proposes the micro payemtn mechanism in which a central server exists, but this
may not be apt or even vcalid for other p2p network like morpheus, gnotella ets which doesnt
have this central server concept.
2. The author doesnt clrify the validity of the formula he uses for the calculation of FT i.e. how
did he arrive at that formula and he dint even mention from where did he adopt that formula
3. The author says that global micro payment sums upto zero but i think he is ignoring or
neglecting the cost of calculation and dissipation/collection of payments.
4. I did not understand how does flat fee also gives rise to FREE RIDER problem (section 2.4)
5. How might be the equilibria change when considering the risk averse agents in micro payment
mechanism ( the authors doesnt provide a formal
6. I am particulalrly interested in seeing the effect of BW variance with equilibria. Having
different BW provided by dialup, dsl, cable, T1 lans etc , i think the proportion of users with
each BW class might have some interesting resluts or effects on the equlibria. One instance
where all the users or on T1 lans , then might be the users tend to share everything as it doesnt
incur a noticeable cost for the utility function but the scenario might be completely different for
agents using dialup
This paper is a on of the many papers that explores the idea of file sharing and its problems on
The possible points of discussion can be:
1. What if the files that are shared on the internet are quantized in terms of quality if the files.
Here the the term file is not limited to just files but even softwares and utilities that can be shared
legally. So the user who has a document which is relatively important and more demanding then
he should be awarded more than regular points .This will cause the real users to share more files.
Similarly users can be charged more for using important document.
2.Also the the award for uploading should be greater than downloading to create an incentive
amongst users to upload files.
3.The users can also be awarded on the fact that they are allowing other users more resources for
file sharing like BW, Disk Space.
4.The identity of the users if kept protected and only known to the centralised server will prevent
both file sharing among friends and also the users remains assured of any legal/illegal licensing
Lastly most of the users are selfish and not altruistic and open the file sharing program when
they want to download stuff so pricing mechanism should be introduced at the centralised server
that ensures users with altrustic tendencies to be be benefitted and other selfish users to pay for
what they are getting.
1. What will happen if we remove the assumption of central servers keeping track of file
downloads/uploads? In the absence of a central authority, the fallback is a community-based
reputation scheme. In this scheme, users obtain reputations positive/negative according to their
sharing. This reputation information can be disseminated to other peers. Can this work?
2. I think the assumption in section 4 f^AD(1) < alpha is not enough. It should be f^AD(i) <
alpha * i. Well, i think it is Ok because they said that f^AD has a minimum value of zero and is
linear. Also, in the same section, FT = alpha(2-...) the denominator should be 2n-4+s.
3. In section 5.2, they reward sharing based on the amount shared not the number of uploads, but
in the analysis they calculate the expected number of uploads?
1. In this game it is hard for the clients to decide which file to put for upload. Because of this
clients may want to maximize their utility but they don't know how. One suggestion is to let the
server direct clients for which files to put for upload. Does this suggestion affect the analysis of
2. In the Micro-payment model, clients are charged on number of files downloaded and
uploaded, consequently, clients may try to chose small files for upload, how this affect the utility
3. In the learning model, they assume the parameters were chosen at random. However some
parameters may have unrealistic "or extreme" values. Assume we choose some fixed values does
this will affect the learning algorithm? Is it possible that we might not converge to equilibrium?
1. The assumption that spoofing and spamming is prevented for this mechanism to work is too
difficult to be ensured.
2. Similar to the fact that rare and easily available files should have different rewards , sharing
files with false , incomplete or corrupt data should be penalized in some way.And many such
factors will come into the payment function. Problems like incomplete songs about which you
come to know only after downloading it , need to be addressed by the mechanism.
3. The problem here is providing incentives for sharing a file.There can be many ways to
formulate a pricing mechanism for the agents.For example in the point based mechanism where
the user collects points for sharing and for downloading , we could have a situation where one
who shares gets more points than one who downloads. How does it work in such a case?
4. Also, how this rewarding can be extended to distributed environment.Napster has a centralised
1. Why do altruistic agents exist in Napster (without pricing mechanism design)? I can't see any
incentive of these agents.
2. The point based mechanisms apply the "contributions to the network" concept in Napster free
rider problems. How can we apply this concept to the internet? ??
3. What is the basic concept of Q-learning? Is it like the concept of Nural network?
4. What is temporal difference Q learning?
5.Why is the utility model a logarithmic function?
1. just sharing the part of file is useful to get the some idea of the file? i,e, the portion of the file
is not enough to know some idea of the file.
2. If you decide to charge the money for the file sharing, who should be responsibe for the wrong
3.Sharing the file when the offer's computer is idle is realistic? i,e, when I need some file, I can't
wait until the offer's computer is idle.