Download presentation

Presentation is loading. Please wait.

Published byZavier Tindell Modified about 1 year ago

1
1 Regret-based Incremental Partial Revelation Mechanism Design Nathanaël Hyafil, Craig Boutilier AAAI 2006 Department of Computer Science University of Toronto

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

3
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
4 “Computational” Mechanism Design The interesting questions: what preference info is relevant to the task at hand? when is the elicitation effort worth the improvement it offers in terms of decision quality? how to deal with incentives ?

5
5 Overview Mechanism Design Background Incremental Partial Revelation Mechanism Regret-based iPRMs Experimental results Conclusion / Future Work

6
6 Basic Social Choice Setup Choice of x from outcomes X Agents 1..n: type t i T i and valuation v i (x, t i ) Type vectors: t T Goal: implement social choice function f: T X e.g., social welfare SW(x,t) = v i (x, t i ) Quasi-linear utility: u i (x, i,t i ) = v i (x, t i ) - i Our focus: social welfare maximization

7
7 Basic Mechanism Design A mechanism m consists of three components: actions A i allocation function O: A X payment functions p i : A R Mechanism is incentive compatible: In equilibrium, agents reveal truthfully Ex-post IC Assume others tell the truth and agent i knows the others’ types Then agent i should tell the truth

8
8 Properties Mechanism is efficient: maximizes social welfare given reported types: -efficient: within of optimal social welfare Ex post individually rational: no agent can lose by participating -IR: can lose at most

9
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
10 Cost of Full Revelation Communication costs Computation costs Cognitive costs Privacy costs INTRACTABLE! Partial revelation?

11
11 Partial Revelation Full revelation: Not always necessary for optimal decision When necessary, not always worth the costs Partial revelation: Elicit just enough to make optimal decision Trade-off elicitation costs with decision quality Can we maintain incentives?

12
12 Existing Work on Partial Revelation [Conen,Hudson,Sandholm, Parkes, Nisan&Segal, Blumrosen&Nisan] Most Work: require enough revelation to determine optimal allocation and VCG payments hence can’t offer savings in general [Nisan&Segal05] Exception: Priority games [Blumrosen&Nisan 02] specific settings (1-item, combinatorial auctions)

13
13 Overview Mechanism Design Background Incremental Partial Revelation Mechanism (iPRM) Regret-based iPRMs Experimental results Conclusion / Future Work

14
14 Incremental Partial Revelation Mechanisms (iPRMs) iPRM interacts with agents: set of queries Q i ( e.g. standard gamble:“v( car ) >5?”) response r R i (q i ) interpreted as partial type i (r) T i (e.g. bounds on each parameter) Formal Model (see paper)

15
15 iPRMs Goal: Trade-off quality of alloc. with revelation costs Maintain acceptable incentives properties At each step, given , choose between: Terminating (which allocation?) Eliciting (which query?)

16
16 Minimax Regret: Utility Uncertainty Regret : Max regret of x given : MMR-optimal allocation: x* = arg min x MR(x, )

17
17 Overview Mechanism Design Background Incremental Partial Revelation Mechanism Regret-based iPRMs Experimental results Conclusion / Future Work

18
18 Regret-based Elicitation Find query to reduce MMR level? Several heuristics proposed for preference elicitation. We adapt Current Solution Strategy (CSS) Focus elicitation on allocations involved in regret

19
19 Allocation Elicitation Proposed allocation elicitation algorithm Using SW-regret computation and elicitation See paper for details Allocation elicitation phase terminates with -efficient allocation Partial type

20
20 Incentive Properties Let mechanism M = (x*, p i T ), with -efficient allocation function x* payments: p i T (x* ; ) = max t p i VCG (x* ; t) Theorem 1: M is -efficient, - ex post IR, - ex post IC = + ( ) ( ): bound on payment uncertainty

21
21 Approximate Incentives : bound on utility gain But gain from manipulation outweighed by costs of manipulation don’t know types of others must simulate mechanism Formal, approximate IC practical, exact IC

22
22 2 Phase Approach Bound on manipulability: + ( ) : not a priori If ( ) too large: Elicit to reduce payment uncertainty Payment elicitation strategy: based on CSS (P-CSS) Terminates with a priori bounds ( + ) -IC -IR, -efficiency

23
23 Direct Optimization Causes of manipulability: efficiency loss + payment uncertainty MMR w.r.t. SW only accounts for efficiency loss Should minimize global worst-case manipulability: u(best lie) - u(truth) efficiency loss bounded by worst-case manipulability Formulate as regret optimization and elicitation ask queries that directly reduce global manipulability

24
24 Single Phase Approach Theorem 2: For M = (x*, p i T ), If =max worst case manipulability Then M is -efficient - ex post IC - ex post IR

25
25 Overview Mechanism Design Background Incremental Partial Revelation Mechanism Regret-based iPRMs Experimental results Conclusion / Future Work

26
26 Elicitation Strategies Two Phase (2P): SW loss and payment uncertainty for elicitation and decisions Two Phase ( 2P): SW loss and payment uncertainty for elicitation Manipulability for decisions Common-Hybrid (CH): Manipulability for elicitation and decisions Myopically Optimal (MY): Simulate all queries, ask best

27
27 Test Domains Car Rental Problem: 1 client, 2 dealers Car: 8 attributes, 2-9 values, ~12000 cars factored valuation/costs: 13 factors, size 1-4 Total 825 parameters Small Random Problems: supplier-selection, 1 buyer, 2 sellers 81 parameters

28
28 Results: Car Rental Initial regret: 99% of opt SW Zero-regret: 71/77 queries Avg remaining uncertainty: 92% vs 64% at zero-manipulability Avg nb params queried: 8% relevant parameters reduces revelation improves decision quality

29
29 Results: Random Problems

30
30 Conclusion Theoretical model for iPRMs Class of iPRMs with approximate incentives Key point: Approximation trade off cost vs. quality Formal, approximate IC practical, exact IC Applicable to general mechanism design Empirically very effective

31
31 Current + Future Work More heuristics + test domains Formal model manipulation and revelation costs formal, exact IC explicit revelation/quality trade-off Sequentially optimal elicitation One-shot partial revelation mechanisms “Mechanism Design with Partial Revelation” draft 2006

32
32 Questions?

Similar presentations

© 2017 SlidePlayer.com Inc.

All rights reserved.

Ads by Google