Survey of Decision Theory - PowerPoint PPT Presentation

1 / 21
About This Presentation
Title:

Survey of Decision Theory

Description:

Each agent wants to maximize his own benefit. Incentive to cheat (lie about your true preferences) ... Could one query engine benefit from the final output? ... – PowerPoint PPT presentation

Number of Views:59
Avg rating:3.0/5.0
Slides: 22
Provided by: larse
Category:

less

Transcript and Presenter's Notes

Title: Survey of Decision Theory


1
Survey of Decision Theory
  • in the Context of Top-k Querying

2
Motivation for Decision Theory
  • Separate agents (not necessarily computers)
  • All agents work together to make a decision that
    affects everyone
  • Each agent has a utility for each possible
    outcome
  • Each agent wants to maximize his own benefit
  • Incentive to cheat (lie about your true
    preferences)

3
Applicability to Top-k
  • Separate attributes
  • Probably all computers, but
  • Not necessarily the same computer
  • Different scoring schemes
  • Need to derive a single scoring function from all
    agents
  • Utility?
  • Certainly from the users perspective
  • Accurate solution
  • Cost of computing accurate solution?
  • Could one query engine benefit from the final
    output?

4
Approaches to Decision Theory
  • Voting
  • Game Theory (Nash Equilibrium, Probability)
  • Implies some sort of utility function to the user
    for each outcome (i.e. each possible top-k
    ordering?)
  • Auctions
  • Bargaining
  • Both imply some sort of utility to each ranking
    agent for each outcome
  • Contracts, coalitions, ?

5
Voting
  • Combine preferences in a way that makes sense
    and is fair to all agents
  • Not all agents will get their preferred outcome
  • Example 2000 USA Presidential Election
  • Normally we select Top-1 over all voters
    preferences
  • Well extend it to Top-n for n candidates

6
Plurality Protocol
  • Scores
  • Bush 3
  • Gore 2
  • Nader 1
  • Buchanan 1
  • Or, remove winner and repeat for 2nd place
  • Nader 3
  • Gore 2
  • Buchanan 2

7
Plurality Protocol Incentive to Cheat
  • Scores
  • Gore 4
  • Bush 3
  • Nader 0
  • Buchanan 0

8
Plurality Protocol
  • Advantages
  • Easy to compute
  • Easy to extend to top-k
  • Disadvantages
  • Incentive to cheat
  • Winner may not rank highly in a majority

9
Binary Protocol
  • Since more people preferred Gore to Bush,
    shouldnt Gore be ranked higher overall than
    Bush?
  • Problem More people preferred Nader to Gore,
    shouldnt Nader be ranked higher?

10
Binary Protocol
  • Ordering 1
  • Bush gt Nader 4/7
  • Bush gt Buchanan 6/7
  • Gore gt Bush 4/7
  • winner Gore
  • Ordering 2
  • Nader gt Gore 5/7
  • Nader gt Buchanan 4/7
  • Bush gt Nader 4/7
  • winner Bush

11
Binary Protocol
  • Advantages
  • Winner must be preferred by the majority in some
    sense
  • Disadvantages
  • Can be non-deterministic, or at least has
    multiple equally valid outcomes
  • Can lead to non-Pareto-efficient solution
  • c gt d gt b gt a a gt c gt d gt b b gt a gt c gt d
  • a wins over c, b wins over a, d wins over b,
    winner is d
  • c is universally preferred over d

12
Arrows Impossibility Theorem
  • All possible preferences must yield a valid
    overall ranking
  • The outcome must be a total ordering
  • Defined for every pair
  • Asymmetric
  • Transitive
  • Pareto efficiency if everyone prefers a to b,
    then the outcome should rank a higher than b
  • Independent of irrelevant alternatives
  • No dictator outcome cannot follow a single
    voters preference regardless of the other voters

13
Borda Protocol
  • Scores
  • Bush 20
  • Gore 17
  • Nader 19
  • Buchanan 14
  • Outcome Bush gt Nader gt Gore gt Buchanan

14
Borda Protocol Irrelevant Alternatives
  • Scores
  • Bush 14
  • Gore 13
  • Nader 15
  • Outcome Nader gt Bush gt Gore

15
Dictator Protocol
Always follow x1! Outcome Bush gt Buchanan gt
Nader gt Gore Or, pick a dictator at
random Outcome x7 (Buchanan gt Nader gt Gore gt
Bush) Not useful for Top-k. Or is it?
16
Dictator Protocol
  • Advantages
  • No incentive to cheat
  • Satisfies all other criteria for Arrows theorem
  • Cost to compute is cheap!!!!
  • If most people prefer x, x has a good chance of
    ranking highly (Random dictator only)
  • Disadvantages
  • Doesnt take other rankings into account
  • Winner may not be ranked highly by other agents

17
Plausibility Protocol
Plausibility of pairs Bush gt Buchanan 6 Nader
gt Gore 5 Bush gt Nader, Gore gt Bush, Nader gt
Buchanan, Gore gt Buchanan 4 Nader gt Bush,
etc 3 Gore gt Nader 2 Buchanan gt Bush 1
18
Plausibility Protocol
  • Plausibility of pairs
  • Bush gt Buchanan 6
  • Nader gt Gore 5
  • Bush gt Nader, Gore gt Bush, Nader gt Buchanan,
  • Gore gt Buchanan 4
  • Pick Bush gt Buchanan and Nader gt Gore.
  • Pick the next preference as the one that
    eliminates the least of the other possibilities
  • Bush gt Nader and Gore gt Bush eliminate each
    other, the other two eliminate nothing

19
Plausibility Protocol
Bush
Buchanan
Gore
Nader
  • Bush gt Nader, Gore gt Bush 4
  • 3
  • Continue until total ordering is achieved.
    Break ties randomly.
  • Two possible outcomes for this scenario

Bush
Gore
Buchanan
Nader
Bush
Gore
Buchanan
Nader
20
Plausibility Protocol
  • Advantages
  • Intuitive idea that the more a pairwise
    preference holds, the more likely it holds in the
    outcome
  • Independent of irrelevant alternatives (?)
  • Remove Buchanan, and the remaining preferences
    still have the same ranking
  • Arrows Impossibility Theorem?
  • Incentive to cheat?
  • Disadvantages
  • Costly to compute
  • Can be non-deterministic

21
How Does This Fit In?
  • Kießling paper Foundations of Preferences in
    Database Systems
  • More about specifying preferences than decision
    theory
  • All the papers that describe aggregation
    functions
  • Scores on each item form a total ordering
  • One query engine can dominate the final outcome
  • No control over scoring functions
  • One or more query engines has some vested
    interest in the final ranking? (e.g. Google gets
    some payola from Amazon.com)
  • Perhaps we want to minimize the impact from
    artificially inflated rankings
Write a Comment
User Comments (0)
About PowerShow.com