Katrin Meder

1
Katrin Meder
Presentation, June 25th

• Golle, Leyton-Brown, Mironov Incentives for
  Sharing in Peer-to-Peer Networks

2
Agenda

• Paper by Golle et al. Theory Results
• Analysis of weak points / critical questioning
• Experimental approach to test and to verify the
  assumptions
• Conclusion

3
The paper by Golle et al. describes

• A variety of architectural designs of P2P systems
  in search of an optimal way to overcome the
  free-rider problem using incentives
• Analyses game theoretic equilibria under certain
  assumptions and under several payment mechanisms

4
Problem Definition

• Use of Napster as a concrete example (centralized
  P2P system)
• maintains DB of files currently available
• connects download requests with available clients
• monitors all exchanges
• able to determine identity of all files provided
  by users

5
Problem Definition

• Usage divided into time periods of equal duration
  (t)
• n agents each having two actions available in
  each time period

• 1. Sharing
• ?0 - no sharing
• ?1 - moderate sharing
• ?2 - heavy sharing

• 2. Downloading
• ?0 - no downloads
• ?1 - moderate downloads
• ?2 - heavy downloads

6
Problem Definition

• Agent ais strategy in t
• pure strategy or
• mixed strategy
• Agents strategy leads to a certain expected
  utility

7
Agent Utility

• Function describes his preferences for different
  outcomes

• AD - Amount to Download ()
• NV - Network Variety ()
• AL - Altruism ()
• DS - Disk Space Used (-)
• BW - Bandwidth Used (-)
• FT - Financial Transfer

8
Agent Utility

• must be monotonically
  increasing
• are constant for values
  greater than and respectively
• must be monotonically decreasing
• take value -? for all values
  greater than k3 and k4 respectively

9
Experimental Testing of the utility function

• Create a P2P network in the lab,
• Users play t periods of time
• For each time period, they are able to set
  preferences for several choice problems along a
  continuous scale (e.g. 0100)
• Over time, peers will experience the consequences
  of their decisions and will reach a preference
  equilibrium

10
Experimental Testing of the utility function

• Parameters to be examined
• long search lists requiring longer search time
  vs. short lists in little time speed, AD, NV
• providing prioritized friends with own bandwidth
  vs. randomly providing bandwidth BW
• anonymity vs. being known and be able to gain
  reputation
• preferring high quality of files with willingness
  to pay for them vs. random quality for free
  Quality
• To be observed
• decreasing number of downloads over time (less
  disk space) vs. continuous downloading with
  burning and deletion of files in between DS

11
Experimental Testing of the utility function

• Aggregated equilibria of all peers represent
  overall utility - illustrates what peers want and
  expect from a P2P system
• Discussion
• How to measure altruism?
• What settings do we expect to prevail?
• Are there levels of satiation for AD and AL? -
  Hypothesis Utility of AD is not constant

12
Analysis of Equilibria

• Non-monetary Situation
• 1. Excluding altruism
• 2. Including altruism
• Monetary Mechanisms
• 3. Micro-payment mechanisms
• 4. Quantized micro-payment mechanisms
• Point-based Mechanisms
• 5. Micro-payments in points
• 6. Rewards for sharing

13
1st Game Setup

• No financial transfers, disregarding altruism
• Weak equilibrium
• Agents dont share because that decreases their
  utility
• Since no one shares NV0

14
2nd Game Setup

• Altruism
• Service free of charge
• Sense of community among users
• Modest disincentive for non-contribution since
  Napster shares all songs s.o. downloaded
• Two types of agents
• Altruistic agents dominant strategy
• Non-altruistic agents strategy

15
2nd Game Setup

• Describes current state of affairs
• All agents are unrestrained in their downloads
• Some agents are sufficiently altruistic to share
  while others share nothing
• Implies FREE RIDER PROBLEM
• next game setup Imposition of financial
  transfers in order to overcome the free rider
  problem

16
Special Flat rates

• Fixed fee per time period
• covers royalty costs or the overhead involved
• No impact on the equilibrium, unrelated to the
  agents behavior - implies FREE RIDER PROBLEM
• Only the magnitude of the flat rate decides
  initially whether agents participate or not, but
  irrelevant for further behavior

17
Discussion / Experiment

• Is the existence of flat rates really unrelated
  to agents behavior? Are users expectations
  still the same?
• Experiment Compare several equally structured
  P2P architectures
• keep external conditions as similar as possible
  (number of peers, types of files, start date,
  etc.)
• only vary the flat rate across networks -
  Threshold?
• allow agents a period of free testing, make them
  pay afterwards
• control the number of files made available by
  different peers
• Does the flat rate paid correlate with
  expectations of the network (speed, data
  quality...)? - questionnaire

18
Micro-Payment Mechanisms

• Concept of penalizing downloads and rewarding
  uploads

19
3rd Game Setup

• Server processes all download requests and keeps
  track of the number...
• of files downloaded ?
• of files uploaded ?
• At the end of t, users are charged
• Global sum of all micro-payments is zero

20
3rd Game Setup

• Server matches downloaders at random with shared
  units
• ? represents costs per net unit downloaded
• Assumptions
• incentive to download as much as
  possible
• Strong equilibrium
• Overcomes Free Riding

21
Problem

• Micro-payment mechanisms require centralization
  of the network with a higher authority that keeps
  track of all exchanges and does the aggregate
  billing
• Discussion Unrealistic for decentralized
  systems?!
• Higher authority has to be paid - requires the
  following inequality
  costs for downloads gt rewards for
  uploads
• Disincentive for use of this network - more
  difficult for some individuals to profit from
  providing many uploads

22
4th Game Setup

• Users dislike micro-payments and prefer flat
  pricing plans Payment for downloads in blocks of
  b files
• Advantage no mental decision costs associated
  with per-download pricing
• b1 - original micro-payment mechanism
• Irrational for agents to download less than a
  multiple of b
• May be attacked by a coalitions of friends that
  use their zero-margin-cost downloads for
  generating money for themselves

23
4th Game Setup - Modifications

• 1. Server provides list of all users serving the
  requested files but without their identities

• 2. Server provides a random subset of all users
  serving the requested files

• Doesnt make zero-margin-cost downloads to
  friends impossible (e.g. rare files)
• Possible solution Treat rare files differently!

24
4th Game Setup - Treatment of rare files

• Idea No credit to users serving rare files
• Problem Strong disincentive for introducing new
  files into the system
• Solution Server observes exchange of rare files.
  If number of downloads breaks a threshold of
  frequency (minimum level of popularity), the user
  is retroactively rewarded.

25
Problem Treatment of rate files

• Still significant disincentive for providing old
  and rare files that wont become popular anymore
• Experiment that traces availability of rare
  files
• pricing of downloads in blocks of b files
• pricing for serving files on an individual basis

• 1. Dont reward uploads that fall below a certain
  threshold of frequency per time period

• 2. Reward every upload equally per time period

• Hypothesis First example provides significantly
  less old, rare files (only from idealists) - less
  network variety - less utility for everyone

26
Problem Treatment of rate files

• Discussion
• Would people prefer micro-payment mechanisms with
  a larger network variety in comparison to
  flat-pricing plans that dont reward rare files
  (at once)?
• How do you want to determine the threshold for
  retroactive payments of new files?

27
Point-Based Mechanisms

• Buy points with money or with contribution to the
  network
• Points are not redeemable for money
• In repeated games, agents wont have incentives
  to accumulate more points than they spend

28
5th Game Setup

• Payment on a per file basis (like 3rd game)
• downloading one file costs 1 point
• uploading one file earns 1 point
• Fixed amount of money for a block of b points
  (like 4th game)
• Not subject to zero-margin-cost downloads

29
5th Game Setup

• Strong equilibrium
• Problem Desire for downloads and uploads are
  connected
• a desire for downloads of ?1 leads to the
  following equilibrium

30
Problems

• People that mainly provide files to the network
  without downloading very much, have no incentive
  to provide more since points are not redeemable
  for money
• Discussion Who receives the payments?
• Developers of the network?
  Only receive money when new
  users want to join or when existing users want to
  download more than than they make available.
• Realistic?

31
Problems

• Inflation
• Prices for a block of b points should regularly
  be adjusted to price fluctuations. (Weekly,
  monthly, yearly...?)
• If prices for points are adjusted, points per
  song remain constant
• Discussion

32
6th Game Setup

• Agents are rewarded for sharing - in proportion
  to the amount of material they make available
• Payment proportional to
• M(t) - amount of MB available for download in t
• Costs for downloading cm points
• m - size of file in MB
• c - constant parameter intuition number of
  hours the file must be shared in order to waive
  the costs of the download
• Costs for 1 point ?

33
6th Game Setup

• Assumptions
• All files have the same size (1 MB)
• parameter c1
• Moderate sharing (?1) - 1 point,
  sharing all files (?2) - 2 points

34
6th Game Setup

• Strict equilibrium played by
  n-1 agents
• Agent ai plays
• no way to cooperate and make money by downloading
  from each other

35
6th Game Setup

• But is not a unique equilibrium
• Another possible equilibrium is
  played by n-1 agents, thus agent n will also not
  share any files (avoid negative utility from
  serving all requests)
• Discussion Which equilibrium is more likely and
  more realistic?

36
6th Game Setup

• Are incentives for sharing files altered due to
  negative utility for the consumption of
  bandwidth?
• 1. Users may make their files available at low
  usage times only
• Solution rewards based on (expected?) download
  demand
• scaling factor
  proportional to (expected?) demand
• 2. Users may only offer unpopular files
• Solution see 1.

37
Problem Centralized vs. a Decentralized System

• E.g. Napster vs. Kazaa
• Revision of assumptions - Discussion
• Are AD, NV, BW and DS independent of
  decentralization?
• same result for Game Setup 1 without altruism

38
Problem Centralized vs. a Decentralized System

• AL in Game Setup 2 - Discussion

• 1st Hypothesis There is more altruism in
  centralized systems because
• Centralized systems represent a central meeting
  point, create a better sense of community
• Represents a central authority that keeps track
  of the files being exchanged - you might feel
  observed (peer pressure)

• 2nd Hypothesis There is less altruism in
  centralized systems because
• Free riding becomes easier since most of the
  peers are unknown to you and expect nothing from
  you
• e.g. at Kazaa you download from several peers
  simultaneously - discourages efforts of personal
  contact

39
Problem Centralized vs. a Decentralized System

• Experiment Compare the level of altruism in
  centralized and decentralized networks on a large
  scale
• Keep external conditions as similar as possible
  (number of peers, types of files, start date,
  etc.)
• Discussion Are users actually aware of the
  network architecture? Do they care?
• Interesting parameters Indicators for altruism?
• number of files made available by different peers
• engagement in personal contact
• Correlation between both? (Hypothesis positive)

40
Problem Centralized vs. a Decentralized System

• Game Setup 2 should yield the same equilibrium
  outcome in a decentralized system as in the
  described central system
• coexistence of altruistic and non-altruistic
  agents

41
Problem Centralized vs. a Decentralized System

• Game Setup 3 4 Micro-payment mechanisms
• Not expected to work in decentralized systems -
  represents a strong disincentive for the use of
  the network
• Could be proven by an experiment

42
Problem Centralized vs. a Decentralized System

• Experiment Based on a small P2P network
• Ask users to keep track of their files uploaded
  and charge the corresponding downloaders a fixed
  price per file downloaded
• Introduction of - Billing Efforts -
  disutility perceived from engaging in private
  billing
• Hypothesis - Agents will
  give up charging money based on free downloads in
  return
• Return to equilibrium from Game Setup 2

43
Non-monetary Incentives

• Reputation - How to be created?
• In small networks, reputation among peers B-i (1
  ? i ? n) is created from seeing / downloading
  what A has available and from telling other peers
  C-i (1 ? i ? m)
• Possibility for larger networks Central
  Evaluation System (e.g. see Ebay) - peers B-i can
  evaluate the quality and availability of files
  downloaded from A in order to create trust or
  distrust among peers C-i
• Experiment
• Does reputation help to overcome free riding?
• How to assign download requests to providers with
  reputation?

44
Non-monetary Incentives

• 1. Experiment
• small P2P network (about 10 agents)
• peers can evaluate the files offered by other
  peers (file source) and their trustworthiness
  after a download with regards to quality and
  availability in a central rating system
• before downloading a file, peers can decide from
  a listing of trustworthiness whether they want to
  choose that peer
• loss of reputation only when peer doesnt share
  any more

45
Non-monetary Incentives

• ...
• problem for new peers that are willing to share
  in order to create reputation everyone may
  prefer established peers only
• Hypothesis for an equilibrium
• few sharing peers
• many downloading peers
• Problem of Free Riders

46
Non-monetary Incentives

• 2. Experiment
• Establishment of reputation in the same central
  way
• Server connects download requests after throwing
  a dice - the higher the reputation of the
  provider, the more probable it is that the
  download is allowed
• Makes it slightly easier for new peers to gain
  reputation from just having popular files
  available and thus to get into the trade (still
  time-consuming)

47
Conclusion

• Applying no incentives at all leads to the
  classical free rider problem
• Monetary incentives (esp. micro-payment
  mechanisms) may lead to theoretical equilibria
  that overcome free riding
• But systems based on monetary incentives may lead
  to even less altruism in reality
• Is reputation an appropriate incentive?

48
The End

• Thanks for your attention.