1 / 31

New Insights on Where to Locate a Library

New Insights on Where to Locate a Library. Ariel D. Procaccia ( Microsoft). Foreword. Best advisor award goes to... Thesis is about computational social choice Approximation Learning Manipulation. BEST ADVISOR. Where to locate a library on a street?.

dillon
Download Presentation

New Insights on Where to Locate a Library

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. New Insights on Where to Locate a Library Ariel D. Procaccia (Microsoft)

  2. Foreword • Best advisor award goes to... • Thesis is about computational social choice • Approximation • Learning • Manipulation BEST ADVISOR

  3. Where to locate a library on a street? • Want to locate a public facility (library, train station) on a street • n agents A, B, C,... report their ideal locations • A mechanism receives the reported locations as input, and returns the location of the facility • Given facility location, cost of an agent = its distance from the facility

  4. Take 1: average • Suppose we have two agents, A and B • Mechanism: take the average • A mechanism is strategyproof if agents can never benefit from lying = the distance from their location cannot decrease by misreporting it • Problem: average is not strategyproof

  5. Take 2: leftmost location • Mechanism: select the leftmost reported location • Mechanism is strategyproof • A mechanism is group strategyproofif a coalition of agents cannot all gain by lying = the distance from at least one member does not decrease • Mechanism is group strategyproof A B B B C D E

  6. Social cost and approximation • Social cost (SC) of facility location = sum of distances to the agents • Leftmost location mechanism can be bad in terms of social cost • One agent at 0, n-1 agents at 1 • Mechanism selects 0, social cost MECH = n1 • Optimal solution selects 1, social cost OPT = 1 • Mechanism gives -approximationif for every instance, MECH/OPT   • Leftmost location mechanism has ratio  n1

  7. Take 3: the median • Mechanism: select the median location • The median is group strategyproof • The median minimizes the social cost B A C D D D E

  8. Facility location on a network • Agents located on a network, represented as graph • Examples: • Network of roads in a city • Telecommunications network: • Line • Hierarchical (tree) • Ring (circle) • Scheduling a daily task: circle A B C

  9. Median on trees • Suppose network is a tree • Mechanism: start from root, move towards majority of agents as long as possible • Mechanism minimizes social cost • Mechanism is (group) strategyproof A A B B D F F C C G E

  10. Strategyproof mechanisms in general networks • Schummer and Vohra [JET 2004] characterized the strategyproof mechanisms on general networks • Corollary: if network contains a cycle, there is no strategyproof mechanism with approx ratio < n1 for SC

  11. A randomized mechanism • A randomized mechanism randomly selects a location • Cost of agent = expected distance from the facility • Social cost = sum of costs = sum of expected distances • Random dictator mechanism: select an agent uniformly and return its location • Theorem: random dictator is a strategyproof (22/n)-approx mechanism for SC on any network

  12. Random dictator is not always group strategyproof • Consider a star with three arms of length one, with three agents at leaves • Cost of each agent = 4/3 • After moving to center, cost of each agent = 1 1/3 A A 1 N 1 1 1/3 1/3 B B C C

  13. Random dictator is sometimes group strategyproof • If the network is a line, random dictator is group strategyproof • Theorem: if the network is a circle, random dictator is group strategyproof

  14. Summary of social cost ? ?

  15. Minimizing the maximum cost • Maximum cost (MC) of facility location = max distance to the agents • Example: facility is a fire station • Optimal solution on a line = average of leftmost and rightmost locations, its max cost = d(A,E)/2 • Mechanism: select A • Mechanism is group strategyproof and gives a 2-approximation to MC • Theorem:There is no deterministic strategyproof mechanism with approx ratio smaller than 2 for MC on a line A C D E B

  16. The Left-Right-Middle Mechanism • Left-Right-Middle (LRM) Mechanism: select leftmost location with prob. ¼, rightmost with prob. ¼, and average with prob. ½ • Approx ratio for MC is[½  (2  OPT) + ½  OPT] / OPT = 3/2 • LRM mechanism is strategyproof • Theorem:LRM Mechanism is group strategyproof • Theorem:There is no randomized strategyproof mechanism with approximation ratio better than 3/2 for MC on a line 2d d B B C A D E 1/2 1/4 1/2 1/4 1/4

  17. Minmax on general networks • Mechanism: choose A • Gives a 2-approximation to the maximum cost • Lower bound of 2 still holds

  18. LRM on a circle • Semicircle like an interval on a line • If all agents are on one semicircle, can apply LRM • Meaningless otherwise 1/4 A B C F D 1/2 E 1/4

  19. Random Midpoint • Look at points antipodal to agents’ locations • Random Midpoint Mechanism: choose midpoint of arc between two antipodal points with prob. proportional to length • Theorem: mechanism is strategyproof • Approx ratio 3/2 if agents are not on one semicircle, but  2 if they are B 1/4 A 3/8 C C A B 3/8

  20. A hybrid mechanism • Mechanism: • If agents are on one semicircle, use LRM Mechanism • If agents are not on one semicircle, use Random Midpoint Mechanism • Theorem: Mechanism is SP and gives 3/2-approximation for MC when network is a circle • Lower bound of 3/2 holds on a circle

  21. A randomized lower bound on trees • Theorem:there is no randomized strategyproof mechanism with approximation ratio better than 2o(1) for MC on trees

  22. Summary of maximum cost ?

  23. Bibliographic notes • Approximate mechanism design without moneyWith Moshe Tennenholtz [EC’09] • Locating a facility on a line • Locating two facilities on a line • Locating one facility on a line when each player controls multiple locations • Strategyproof approximation mechanisms for location on networksWith NogaAlon, Michal Feldman, and Moshe Tennenholtz [under submission] • Locating a facility on a network • Available from Google: Ariel Procaccia

  24. A bit on algorithmic mechanism design • Algorithmic mechanism design (AMD) was introduced by Nisan and Ronen [STOC 1999] • The field deals with designing strategyproof (incentive compatible) approximation mechanisms for game-theoretic versions of optimization problems • All the work in the field considers mechanisms with payments • Money unavailable in many settings

  25. Class 1 Opt SP mechanism with money Problem is intractable Opt SP mech with money + tractable Class 3 No opt SP mech w/o money Class 2 No opt SP mech with money

  26. Approximate mechanism design without money • Can consider computationally tractable optimization problem • Approximation to obtain strategyproofness rather than circumvent computational complexity • Originates from work on incentive compatible regression learning and classification [Dekel+Fischer+P, SODA 08, Meir+P+Rosenschein, AAAI 08, IJCAI 09]

  27. Future work • I Promised “avalanche of challenging directions for future research” • I lied • Generally speaking: • Many technical open questions • Many extensions, can combine extensions • Completely different settings

  28. Thank Y u!

  29. Current work • Agents are vertices in directed graph, score is indegree • Must elect a subset of agents of size k • Objective function: sum of scores of elected agents • Strategy of an agent: outgoing edges • Utility of an agent: 1 if elected, 0 if not

  30. Lower bound of two • Theorem: there is no deterministic strategyproof mechanism with approx ratio smaller than 2 on a line • Suppose mechanism has ratio < 2 • Let A = 0, B = 1; OPT = ½ • Mechanism must locate facility at 0 < x < 1 • Let A = 0, B = x; OPT = x/2 • Mechanism must locate facility at 0 < y < x • B gains by reporting 1 A B B B 0 1

  31. Minmax on general networks • Mechanism: choose A • Gives a 2-approximation to the maximum cost • O = optimal location, X = some agent • d(A,X)  d(A,O) + d(O,X)  2  OPT • Lower bound of 2 still holds

More Related