Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff. Basic Background – What is the problem? ...

59
Requirements for Fair and Robust Voting Systems Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff

Transcript of Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff. Basic Background – What is the problem? ...

Page 1: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Requirements for Fair and Robust Voting Systems

Mark WangJohn Sturm

Sanjeev KulkarniPaul Cuff

Page 2: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Basic Background – What is the problem? Condorcet = IIA Survey Data Pairwise Boundaries = No strategic voting

Outline

Page 3: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

How to make group decisions when you must.◦ Corporations◦ Organizations◦ Academic Departments◦ Politics

Democracy

Page 4: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Majority Decision No real controversy

Three candidates or more --- not so obvious

Only Two Choices

Page 5: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Voters express preferences as a list

Preferential Voting

Page 6: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Vote Profile (entire preferences of population)

◦ Overly simplified for illustration

Voting Example

Page 7: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Vote Profile

Each voter submits only one name on ballot The candidate named the most wins Ann wins with 37% of the vote.*

Plurality

Page 8: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Vote Profile

Rounds of plurality voting, eliminating one loser each round.

Betty loses first, then Carl wins with 63%.

Instant Run-off

Page 9: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Vote Profile

Score given for place on ballot Ann = .74, Betty = 1.3, Carl = .96 Betty wins

Borda

Page 10: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Voting systems can give different outcomes◦ Do they really?

Is there a best system?◦ How do we arrive at it?

Item 2, A Perfect Voting System

Page 11: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Natural Assumptions:1. (Anonymous) Each voter is treated equally2. (Neutral) Each candidate is treated equally3. (Majority) Two candidates = majority4. (Scale invariant – to be defined)

Assumptions

Page 12: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

1. Robust to Candidate adjustments◦ Independence of Irrelevant Alternatives (IIA)

2. Robust to Voter adjustments◦ Immunity to Strategic Voting

Golden Properties

Page 13: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

The voting outcome would not be different if any of the non-winning candidates had not run in the race.

Our Definition of IIA

Page 14: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

One assumption:◦ Given someone’s honest preference order, we can

assume they would vote for the higher ranked candidate in a two-candidate election.

IIA => Winner must win in any subset◦ => winner must win pairwise against everyone◦ => Condorcet winner must exist and win election

Condorcet = IIA

Page 15: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Arrow’s Impossibility Theorem

Condorcet winner may not exist

Problem – IIA is impossible!

Page 16: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Anonymous:◦ Outcome only depends on number of voters who

cast each possible ballot.◦ i.e. ( #(A,B,C), #(A,C,B), … )◦ Equivalent: ( #(A,B,C)/n, #(A,C,B)/n, … ) and n.

Scale Invariant:◦ Only depends on ( #(A,B,C)/n, #(A,C,B)/n, … ) ◦ i.e. Empirical distribution of votes

Graphical Representation

Page 17: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Simplex Representation

(A > B > C)

(A > C > B)

(B > A > C)

(B > C > A)

Vote Profile

Page 18: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Simplex

Page 19: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Some 3-candidate voting systems can be visualized through projections of the simplex

Example: Three dimensional space of pair-wise comparisons◦ Condorcet Regions◦ Many Condorcet Methods◦ Borda

Projections of the Simplex

Page 20: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Condorcet Regions

Page 21: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Condorcet Regions

Page 22: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Condorcet Regions

Page 23: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Borda

Page 24: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Black Method

Page 25: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Baldwin Method

Page 26: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Kemeny-Young Method

Page 27: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

What really happens?◦ In real life, does a Condorcet winner exist or not?

Item 3, Data

Page 28: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Tideman Data (1987, supported by NSF grant)◦ UK Labour Party Leader◦ Ice Skating

American Psychology Association Elections Debian Leader Election (2001-2007) Various City Elections Our collection: Survey responses from 2012

US Presidential Election

Sources of Data

Page 29: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Visualization of Tideman Data

Sets of 3 candidates

Top View Side View

Page 30: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Visualization of Tideman Data

Sets of 3 candidates with enough voters

Top View Side View

Page 31: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Do voters have an incentive to vote strategically?◦ A strategic vote is a vote other than the true

preferences. Does this ever benefit the voter?

Gibbard-Satterthwaite Theorem◦ Yes

Strategic Voting

Page 32: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Immunity to Strategic Voting is a property of the shape of the boundaries.

Some boundaries may not give incentive for strategic voting.

Boundary Property

Page 33: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

BordaAll boundaries incentivize strategic voting

Page 34: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Kemeny-Young Method

Page 35: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

B

Consider a planar boundary

Requirement for Robust Boundary

A

Page 36: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Robust boundaries are based only on pairwise comparisons

Robust Boundaries

Page 37: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Assume anonymity◦ (more than G-S assumes)

Assume at least 3 candidates can win Not possible to partition using only non-

strategic boundaries.

Geometric Proof of G-S Theorem

Page 38: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Condorcet RegionsIn some sense, this is uniquely non-strategic.

Page 39: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Voters are allowed to modify their vote based on feedback◦ The population is sampled◦ One random voter is allowed to change his vote◦ Update the feedback… pick another random voter

Dynamic Setting

Page 40: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

US Presidential Election Survey

Results2012 Poll

Page 41: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

All Results

Page 42: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Google Poll

Page 43: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Mercer County Results

Page 44: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Mercer County

Page 45: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Mechanical Turk Poll

Page 46: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Mechanical Turk

Page 47: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

NYT Blog Poll

Page 48: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Google – Condorcet

Page 49: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Google – Plurality

Page 50: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Google – Plurality

Page 51: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Google – Instant Run-off

Page 52: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Google – Borda

Page 53: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Google – Condorcet(Democrat voters)

Page 54: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Google – Condorcet(Republican voters)

Page 55: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Google – Borda(Republican voters)

Page 56: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Google – Range Voting(Republican voters)

Page 57: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Google – Condorcet(Independent voters)

Page 58: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Google – Borda(Independent voters)

Page 59: Mark Wang John Sturm Sanjeev Kulkarni Paul Cuff.  Basic Background – What is the problem?  Condorcet = IIA  Survey Data  Pairwise Boundaries = No.

Condorcet winner is uniquely:◦ IIA

Whenever possible◦ Robust to Strategic Voting

Within the Condorcet regions

Upon strategy, some methods become Condorcet

Disagreement among voting systems is non-negligible.

Summary