Raphael Eidenbenz ETH Zurich Switzerland Stefan Schmid TU Munich Germany TexPoint fonts used in EMF Read the TexPoint manual before you delete this box A A A A A A A Raphael Eidenbenz GameNets 09 ID: 415194
Download Presentation The PPT/PDF document "1 Combinatorial Agency with Audits" is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
1
Combinatorial Agency with Audits
Raphael EidenbenzETH Zurich, Switzerland
Stefan Schmid
TU Munich, Germany
TexPoint fonts used in EMF.
Read the TexPoint manual before you delete this box.:
A
A
A
A
A
A
ASlide2
Raphael Eidenbenz, GameNets ‘09
2
IntroductionGrid Computing...Distributed project orchestrated by one serverServer distributes tasksAgents compute subtask
Results are sent back to serverServer aggregates result
Server / Principal
AgentsSlide3
Raphael Eidenbenz, GameNets ‘09
3Introduction: Grid Computing
What are an agent‘s incentives?Payment, fame, altruismWhy not cheat and return a random result?Will principal find out?Not reallyIndividual computation is a hidden actionPrincipal can only check whether entire project failed
Server / Principal
AgentsSlide4
Raphael Eidenbenz, GameNets ‘09
4Introduction: Grid Computing
Project failedWho did a bad job?Whom to pay?Maybe project still succeedsif only one agent exerts low effortIf more than 2/3 of the agents exert high effort...Whom to pay?
Server / Principal
AgentsSlide5
Raphael Eidenbenz, GameNets ‘09
5
Binary Combinatorial Agency [Babaioff, Feldman, Nisan 2006]1 principal , n selfish risk-neutral agentsHidden actions={high effort, low effort}High effort subtask succeeds with probability δLow effort
subtask succeeds with probability γ
Combinatorial project success functionAND: success if all subtasks succeed
OR: success if at least one subtask succeedsMAJORITY: success if more than half of the agents succeedPrincipal contracts with agentsIndividual payment pi depending on entire project‘s outcomeAssume Nash equilibrium in the created gameSlide6
Raphael Eidenbenz, GameNets ‘09
6
Results [Babaioff, Feldman, Nisan 2006]AND technologyPrincipal either contracts with all agents or with noneDepending on her valuation vOne transition point where optimal choice changesOR technologyPrincipal contracts with k agents,
0· k
· nWith increasing valuation
v, there are n transition points where the optimal number k increases by 1Slide7
Raphael Eidenbenz, GameNets ‘09
7Combinatorial Agency with Audits
Grid computing: server can recompute a subtaskActions are observable at a certain cost κ.Principal conducts k random audits among the l contracted agentsAgent i is audited with probability Sophisticated contractsIf audited and convicted of low effort ! p
i=0 even if project successful
Server / Principal
Agents
Slide8
Raphael Eidenbenz, GameNets ‘09
8Some Observations
The possibility of auditing can never be detrimentalNash Equilibrium if principal contracts l and audits k agentspayment piprincipal utility uagent utility uiSlide9
Raphael Eidenbenz, GameNets ‘09
9AND-Technology
Project succeeds if all agents succeedδ: agent success probability with high effortγ: agent success probability with low effort
There is one transition point
v
*for v· v*, contract no agentfor v¸
v*, contract with all agents and conduct
k* audits
Transition earlier with the leverage of audits
TheoremSlide10
Raphael Eidenbenz, GameNets ‘09
10AND-Technology (2 Agents): Principal Utility Slide11
Raphael Eidenbenz, GameNets ‘09
11AND-Technology: Benefit from Audits in %Slide12
Raphael Eidenbenz, GameNets ‘09
12OR-Technology
Project succeeds if at least one agent succeedsδ: agent success probability with high effortγ: agent success probability with low effort
There are
n
transition point v1*,v2*, ... ,v
n*
for
v ·
v
1
*
,
contract no agent
for
v
l-1
*
·
v
·
v
l
*
, contract with
l
agents, conduct
k
*
(l)
audits
for
v
¸
v
n
*
, contract with all agents and conduct
k
*
(n)
audits
Conjecture
LemmaSlide13
Raphael Eidenbenz, GameNets ‘09
13OR-Technology (2 Players): Benefit from Audits in %Slide14
Raphael Eidenbenz, GameNets ‘09
14Conclusion
If hidden actions can be revealed at a certain cost, the coordinator may improve cooperation and efficiency in a distributed systemAND technologyGeneral solution to optimally choose pi, l and k One transition point with increasing valuationOR technologyFormula for number of audits to conduct if number of contracts givenPrincipal can find optimal solution in O(n)Probably n transition pointsTransition points occur earlier with the leverage of auditsSlide15
Raphael Eidenbenz, GameNets ‘09
15OutlookTest results in the wild
Accuracy of the model?Does psychological aversion against control come into play?Non-anonymous technologiesWhich set of agents to audit?Solve problem independent of technologyAre there general algorithms to solve the principal‘s optimization problem for arbitrary technologies?What is the complexity?Total rationality unrealistic
Thank
you!Slide16
Raphael Eidenbenz, GameNets ‘09
16Bibliography[Babaioff, Feldman, Nisan 2006]:
Combinatorial Agency. EC 2006.[Babaioff, Feldman, Nisan 2006]: Mixed Strategies in Combinatorial Agency. WINE 2006.[Monderer, Tennenholtz]: k-Implementation. EC 2003.[Enzle, Anderson]: Surveillant Intentions and Intrinsic Motivation. J. Personality and Social Psychology 64, 1993.[Fehr, Klein, Schmidt]: Fairness and Contract Design. Econometrica 75, 2007.Slide17
Raphael Eidenbenz, GameNets ‘09
17Outline
Introduction: Grid ComputingCombinatorial AgencyBinary ModelResults by Babaioff, Feldman, NisanCombinatorial Agency with AuditsFirst FactsAND technologyOR technologyConclusionOutlookSlide18
Raphael Eidenbenz, GameNets ‘09
18Anonymous TechnologiesSuccess function
t depends only on number of agents exerting high efforttm: success probability if m agents exert high effortOptimal paymentsPrincipal utilityOptimal #auditsSlide19
Raphael Eidenbenz, GameNets ‘09
19AND-Technology
Project succeeds if all agents succeedSuccess function tm=δm¢γn-m
There is one transition point v
*for v
· v*, contract no agentfor v¸ v*, contract with all agents and conduct k
* audits
TheoremSlide20
Raphael Eidenbenz, GameNets ‘09
20AND-Technology: Principal Utility Slide21
Raphael Eidenbenz, GameNets ‘09
21MAJORITY TechnologyOptimal payment
wherePrincipal utility