1 / 38

(One-shot) Mechanism Design with Partial Revelation

(One-shot) Mechanism Design with Partial Revelation. Nathana ë l Hyafil, Craig Boutilier IJCAI 2007 Department of Computer Science University of Toronto. $$. $$. $$. $$. $$. $$. $$. $$. Bargaining for a Car. Luggage Capacity? Two Door? Cost? Engine Size? Color? Options?.

josiah
Download Presentation

(One-shot) Mechanism Design with Partial Revelation

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. (One-shot) Mechanism Design with Partial Revelation Nathanaël Hyafil, Craig Boutilier IJCAI 2007 Department of Computer Science University of Toronto

  2. $$ $$ $$ $$ $$ $$ $$ $$ Bargaining for a Car Luggage Capacity? Two Door? Cost? Engine Size? Color? Options?

  3. Mechanism Design • Mechanism design tackles this: • Design rules of game to induce behavior that leads to maximization of some objective(e.g., social welfare, revenue, ...) • Objective value depends on private information held by self-interested agents  Elicitation + Incentives

  4. Partial Revelation Mechanism Design • Problem: • Stating full utility is intractable • Costs: communication, computational… • Partial Revelation: • what preference info is relevant to decision? • when is the elicitation cost worth the improvement in decision quality? • how to deal with incentives ?

  5. Overview • Mechanism Design Background • Partial Revelation Mechanisms (PRM) • Regret-based PRMs • Partition Optimization • Experimental Results

  6. Basic Social Choice Setup • Choice of x from outcomes X(e.g. cars) • Agents 1..n: typetiTi and valuationvi(x, ti) • Type vectors: tT • Goal: implement social choice functionf: T  X • e.g., social welfare SW(x,t) =  vi(x, ti) • Quasi-linear utility: • ui(x, i ,ti ) = vi(x, ti ) - i • Our focus: social welfare maximization

  7. Basic Mechanism Design • A direct mechanismM consists of three components: • types Ti • allocation function m: T X • payment functions pi : T R • Mechanism is incentive compatible: (IC) • In equilibrium, agents reveal truthfully • Dominant Strategy IC • Regardless of what others report, agent i should always tell the truth

  8. Properties • Mechanism is efficient: • maximizes social welfare given reported types: •  -efficient: within  of optimal social welfare • Mechanism is Individually Rational: (IR) • no agent can lose by participating • -IR: can lose at most 

  9. Direct Mechanisms • Revelation principle: focus on direct mechanisms where agents directly and (in eq.) truthfully reveal their full types • For example, Groves scheme (e.g., VCG): • choose efficient allocation and use payment function: • incentive compatible in dominant strategies • efficient, individually rational

  10. Cost of Full Revelation • Communication costs • Computation costs • Cognitive costs • Privacy costs INTRACTABLE! Partial revelation?

  11. Existing Work on Partial Revelation [Conen,Hudson,Sandholm, Parkes, Nisan&Segal, Blumrosen&Nisan, …] • Full revelation not always necessary for optimal decision(though worst-case is exponential: [Nisan&Segal05]) • Most Approaches: • require enough revelation for optimal VCG outcome • sequential, not one-shot / specific settings (1-item,CAs) • BUT: optimal decision not always worth the costs • Partial revelation:Trade-off elicitation costs with decision quality • e.g. Priority games [Blumrosen&Nisan 02] • Can we maintain incentives?

  12. Overview • Mechanism Design Background • Partial Revelation Mechanisms (PRM) • Regret-based PRMs • Partition Optimization • Experimental Results

  13. Partial Revelation Mechanisms • A partial type is any subset i Ti • e.g. v(red,2doors)  [50,75], etc… • A one-shot (direct) partial revelation mechanism • set iof partial types, i . (typically partition, not required) • m:   X, chooses allocation m() • pi:   R, sets payment pi() • A truthful strategy: report i s.t. ti i • Goal: • Tradeoff “quality” with revelation/communication costs • maintain appropriate incentives

  14. Partial Revelation MD: Negative Results • Partial revelation  can’t generally maximize SW • must allocate under type uncertainty • Roberts: Dominant-IC  (affine) SW maximizer, • Partial revelation  no Dominant-IC • What are some solutions? • relax solution concept to BNE / Ex-Post • relax solution concept to approximate dominant-IC

  15. Partial Revelation MD: Negative Results • Avoid Roberts by relaxing solution concept? • Bayes-Nash Equilibrium • Theorem: • Bayes-Nash IC PRM with certain form of partitions  Trivial mechanism • Consequences: • max expected SW = same as best trivial • max expected revenue = same as best trivial • “Useless” • Ex-Post Equilibrium: Same

  16. Approximate Incentives •  : bound on utility gain • difference b/w u(best lie) and u(truth) • Considerable costs of manipulation: • Uncertainty over others’ types • Valuation + computational costs • If  is small enoughFormal, approximate IC  practical, exact IC

  17. Overview • Mechanism Design Background • Partial Revelation Mechanisms (PRM) • Regret-based PRMs • Partition Optimization • Experimental Results

  18. Regret-based PRMs • In any PRM, how is allocation to be chosen? • x*() is minimax-regret optimal decision for  • A regret-based PRM:m()=x*() for all  

  19. Regret-based PRMs: Efficiency • Obs: If MR(x*(),)   for all , then regret-based PRM m is -efficient for truthtelling agents. • We can tradeoff efficiency for elicitation effort • More elicitation effort  more refined ’s  smaller  • Incentives?

  20. Regret-based PRMs: Incentives • Can generalize Groves payments • fi (-i): arbitrary type in -i and hi (-i) an arbitrary function of -i • Theorem: Let m be a regret-based PRM with • partial types  and a • partial Groves payment scheme. If MR(x*(),)   for all , then m is -dominant incentive compatible

  21. Approximate Incentives and IR • Can generalize Clark payments to get -IR • A Clark-style regret-based PRM gives • approximate Efficiency • approximate Incentive Compatibility • approximate Individual Rationality • (Increased revenue from flexible payments) • Allows tradeoff “quality” vs revelation costs • as long as we can find a good set of partial types

  22. Overview • Mechanism Design Background • Partial Revelation Mechanisms (PRM) • Regret-based PRMs • Partition Optimization • Experimental Results

  23. (One-shot) Partial Type Optimization • Designing PRM: must pick partial types • we focus on bounds on utility parameters • Use regret-based heuristics to estimate VOI p2 i : p1

  24. (1,… i,…n ) The Mechanism Tree Heuristic: Split 1 Worst-case

  25. (1,… i,…n ) (’1,… i,…n ) (’’1,… i,…n ) The Mechanism Tree Heuristic: Split i Worst-case

  26. (1,… i,…n ) (’1,… i,…n ) (’’1,… i,…n ) (’1,… ’i,… ) (’1,… ’’i,… ) The Mechanism Tree More details necessary to make it tractable

  27. Empirical Results • Negotiation problem • 1 buyer, 1 seller, 4 boolean attributes • valuation/cost given by factored model (GAI) • 16 values/costs specified by 8 parameters • Compare: • uniform partitioning vs. regret-based heuristic • worst-case  and expected  (uniform prior)

  28. Empirical Results • worst  = 90 • 5.5 vs 11 bits • (50% savings) average  = 70 • 6.5 vs 11 bits (40% savings)

  29. Empirical Results • Mechanism accounts for all types • Initial regret: 50-146% of optimal • (depending on actual type vector) • With 11 bits (1.4 bits/param , 0.7 bits/good): • 20-56% of optimal (regret) vs 30-86% (uniform) • 60% reduction of  vs 38%

  30. Contributions • Negative Results • Exact incentives “useless” • Regret-based PRMs • Trade-off “quality” with revelation costs • Partial Types Optimization • Avoid exponential blow-up • Use regret to guide elicitation effectively

  31. Current + Future Work • Sequential PRMs (Hyafil Boutilier AAAI 06) • Formal model manipulation and revelation costs  formal, exact IC  explicit revelation/quality trade-off • Partial Revelation Automated Mech Design • General objective functions • include “execution costs”

  32. QUESTIONS?

  33. Regret-based PRMs: Rationality • Can generalize Clark payments as well • fi (-i): arbitrary type in -I • Thm: Let m be a regret-based PRM with • partial types  and a • partial Clark payment scheme. If MR(x*(),)   for all , then m is -individually rational.

  34. (One-shot) Partial Type Optimization • Designing PRM: must pick partial types • we focus on bounds on utility parameters • A simple greedy approach • Let  be current partial type vectors (initially {T} ) • Let  =(1,… i,…n )   be partial type vector with greatest MMR • Choose agent i and suitable split of partial type i into ’i and ’’i • Replace all [i] by pair of vectors: i ’i ;’’i • Repeat until bound  is acceptable

  35. (1,… i,…n ) (’1,… i,…n ) (’’1,… i,…n ) (’1,… ’i,… ) (’1,… ’’i,… ) (’’1,… ’i,… ) (’’1,… ’’i,… ) The Mechanism Tree Heuristic: Split 1 Heuristic: Split i Worst-case *

  36. A More Refined Approach • Simple model has drawbacks • exponential blowup (“naïve” resolution) • split of i useful in reducing regret in one partial type vector , but is applied at all partial type vectors • Refinement: variable resolution • apply split only at leaves where it is “useful” • Ignore on other leaves • keeps tree from blowing up, saves computation • new splits traded off against “cached” splits

  37. Naïve vs. Variable Resolution p2 p2 p1 p1 i i

  38. Heuristic for Choosing Splits • Adapted from single agent preference elicitation techniques: Current Solution Strategy • Let  be partial type vector with max MR • optimal solution x* regret-maximizing witness xw • intuition: focus on parameters that contribute to regret • reducing u.b. on xw or increasing l.b. on x* helps • But: have to account for both “answers” • Here: also consider second best MR

More Related