Mechanism Design: Online Auction or Packet Scheduling - PowerPoint PPT Presentation

About This Presentation
Title:

Mechanism Design: Online Auction or Packet Scheduling

Description:

Weight currently running agent's value by extra 2d where d is how long it has run for ... v(i) or when there are exactly d(i)-t players that did not drop yet ... – PowerPoint PPT presentation

Number of Views:48
Avg rating:3.0/5.0
Slides: 12
Provided by: erict9
Learn more at: http://www.cse.msu.edu
Category:

less

Transcript and Presenter's Notes

Title: Mechanism Design: Online Auction or Packet Scheduling


1
Mechanism Design Online Auction or Packet
Scheduling
  • Online auction of a reusable good (packet slots)
  • Agents types (arrival, departure, value)
  • Agents can lie about value
  • Agents can lie about arrival departure
  • Restrict to later arrival, earlier departure
  • Goals
  • Maximize value of agents who receive good
  • Maximize revenue generated by auctioneer

2
Reminder of previous results
  • Upper bound of 2
  • Lower bound of f (v5 1)/2 1.618

3
Upper bound of 2
  • Greedy
  • Always send feasible packet with maximum value
  • Greedy is 2-competitive
  • Come up with a 2 packet instance which gives
    lower bound of 2

4
Lower Bound f (v5 1)/2
Figures from Online Scheduling with Partial Job
Values Does Timesharing or Randomization Help?
by Chin and Fung, Algorithmica, 37, 149-164, 2003.
5
Mechanism Design Bounds
  • Agents can lie about values, arrival time, and
    departure time
  • Unbounded
  • Can create 3-competitive mechanism using Set Nash
    concept
  • Agents can lie about values, arrival time, and
    early departure time
  • How could we enforce such a mechanism?
  • Bound of 2 exactly

6
General Lower Bound
  • Lavi and Nisan (SODA 2005)
  • Must have restriction on deadline or else cannot
    guarantee bounded competitive ratio
  • Key observation
  • Consider price pi(b-i) faced by agent i at time 1
  • Suppose v(i) lt pi(b-i). Can agent i ever get
    item i?
  • Suppose v(i) gt pi(b-i) but doesnt win item 1
  • Now have M agents all arrive at time 1 with
    deadlines M and values in the range of (1, 1e).
  • Only one agent wins item in first time slot
  • Optimal allocation is all agents win an item in
    some slot
  • M can be arbitarily large so no bound on
    competitive ratio

7
Restricted lower bound of 2
  • Hajiaghayi, Kleinberg, Mahdian, Parkes (EC 2005)
    and no 2-e mechanism
  • Describe what happens in the following scenarios
  • (1,1,infinity) and (1,2,1) (ar, dep, value).
  • (1,1,1d) and (1,2,1) (what about price?)
  • (1,2,1d) and (1,2,1) and (2,2,infinity)
  • (1,2,1d) and (1,1,1) and (2,2,infinity)
  • (1,2,1d) and (1,1,1)

8
Restricted upper bound of 2
  • Based on greedy 2-competitive algorithm
  • Allocation
  • In each time slot, give item to highest bidder
  • Price computation
  • Second price auction
  • Price can drop in later rounds if it could have
    gotten the item cheaper in a later round
  • Example
  • (1,2,2), (1,1, 2-e), (2,2,1)

9
Variations
  • k copies of each item available in each time slot
  • Basically the same except the k top bidders in
    each time slot get the item
  • Asynchronous time slots
  • Item is needed for 1 unit of time but not all
    arrivals/deadlines are at integer time points
  • 5 competitive mechanism
  • Weight currently running agents value by extra
    2d where d is how long it has run for

10
Set Nash Idea
  • Identify a set of recommended strategies for
    all players
  • Set-Nash Equilibria A best response to all other
    agents playing a recommended strategy is to
    employ some recommended strategy
  • Truthful mechanism set of strategies is 1,
    truthfulness
  • Not as powerful as truthful strategy is best
    response to ANY combination of strategies from
    other agents
  • Any game set could be all strategies and then
    this is trivially true

11
Application
  • Japanese auction tradition incremental auction
    (i.e. bids raise by e until there is a winner)
  • Sequential Japanese auction use Japanese auction
    at each time t
  • Players observe dropouts
  • Myopic strategy
  • Drop out when price reaches v(i) or when there
    are exactly d(i)-t players that did not drop yet
  • Semi-myopic strategy
  • Drop no later than when price reaches v(i) and,
    satisfying first condition, no earlier than when
    only d(i)-t players did not drop yet
  • If all players employ a semi-myopic strategy,
    3-approximation
  • Set-Nash Equilibria The set of semi-myopic
    strategies forms a set Nash equilibrium
Write a Comment
User Comments (0)
About PowerShow.com