voting in parallel universes
play

Voting in Parallel Universes ILLC Workshop on Collective Decision - PowerPoint PPT Presentation

Voting in Parallel Universes ILLC Workshop on Collective Decision Making 2015 Stphane Airiau LAMSADE joint work with Justin Kruger and Jrme Lang LAMSADE - Universit Paris-Dauphine ILLC Workshop on Collective Decision Making 2015 1


  1. Voting in Parallel Universes ILLC Workshop on Collective Decision Making 2015 Stéphane Airiau LAMSADE joint work with Justin Kruger and Jérôme Lang LAMSADE - Université Paris-Dauphine ILLC Workshop on Collective Decision Making 2015 1 Stéphane Airiau (LAMSADE) - Voting in Parallel Universes

  2. Voting There is no perfect voting rule There is no consensus on using a particular rule Ties do occur Some voting rules tend to have a large set of winners. Can we use existing rules to define rules that are more deci- sive and less sensitive to tie-breaking rules? ILLC Workshop on Collective Decision Making 2015 2 Stéphane Airiau (LAMSADE) - Voting in Parallel Universes

  3. Notations N is the set of n voters C is the set of m candidates each voter has a preference ≻ i over the set of candidates. We assume the preference is a linear order over the set of candidates We can also view a linear order as a permutation . So we will write S ( X ) the set of all permutations/linear orders on the set X . a profile is an element of S ( C ) n , i.e. a vector �≻ 1 ,..., ≻ n � Definition ((Irresolute) Social Choice Function) A social choice function is a mapping f : S ( C ) n → 2 C The set f ( �≻ 1 ,..., ≻ n � ) is the set of winners . ILLC Workshop on Collective Decision Making 2015 3 Stéphane Airiau (LAMSADE) - Voting in Parallel Universes

  4. Dealing with Ties breaking ties by breaking anonymity : we break a tie using the preference of a special voter (e.g. the president of a committee breaks the ties, or the oldest) ➫ not all voters are equal breaking ties by breaking neutrality : we break a tie using some relation over the candidates : break the ties in favor of the oldest/yougest candidate or using lexicographic order on their names ➫ not all candidates are equal We will focus on the approach breaking neutrality. Definition (Permutation rule) We call a permutation rule a mapping f : S ( C ) n × S ( C ) → C We can view f as an irresolute voting rule attached with a tie-breaking rule ⊲ . ILLC Workshop on Collective Decision Making 2015 4 Stéphane Airiau (LAMSADE) - Voting in Parallel Universes

  5. new rules: each tie-breaking defines a different universe Given a permutation rule f , we can define two new irresolute voting rules: union rule uf : S ( C ) n → 2 C such that uf ( �≻ 1 ,..., ≻ n � ) = { c ∈ C | ∃ ⊲ ∈ S ( C ) | f ( �≻ 1 ,..., ≻ n � , ⊲ ) = c } This rule selects the candidates that win at least once with a permutation rule. argmax rule af : S ( C ) n → 2 C such that af ( �≻ 1 ,..., ≻ n � ) = max c ∈ C |{ ⊲ ∈ S ( C ) | f ( �≻ 1 ,..., ≻ n � , ⊲ ) = c }| This rule selects the candidates that most often win over all permutation rules. ILLC Workshop on Collective Decision Making 2015 5 Stéphane Airiau (LAMSADE) - Voting in Parallel Universes

  6. a new social decision scheme A Social Decision scheme if a mapping S ( C ) → ∆ ( C ) where ∆ ( C ) denotes the set of all probability distributions over the set of candidates. frequency rule Given a permutation rule f , we can define a new social decision scheme pf : S ( C ) n → 2 C such that pf ( �≻ 1 ,..., ≻ n � )( c ) = |{ ⊲ ∈ S ( C ) | f ( �≻ 1 ,..., ≻ n � , ⊲ ) = c }| n ! ILLC Workshop on Collective Decision Making 2015 6 Stéphane Airiau (LAMSADE) - Voting in Parallel Universes

  7. Case study: Single Transferable Vote (also called Instant Run-Off Voting) STV is an iterative rule that works as follows: at each round, each voter casts a ballot containing its favourite candidate We cound the number of votes for each candidate if a voter obtains a majority: it is elected otherwise we eliminate the candidate with the smallest number of votes and we iterate the process with the reduced set of candidates the process eventually stops as either a candidates gets the majority or because it is the only candidate left ➫ there can be ties between candidates that got the smallest number of votes! ILLC Workshop on Collective Decision Making 2015 7 Stéphane Airiau (LAMSADE) - Voting in Parallel Universes

  8. Example 10 voters named 1, 2, ..., 10 3 candidates a , b and c we note the preference a ≻ b ≻ c as abc number preference of voters 4 a b c 3 b c a 2 c b a 1 c a b a gets 4 votes, b and c are tied with 3 ➫ two universes: one where b is eliminated the other where c is eliminated when b is removed: c wins when c is removed: a and b are tied again! Conitzer, Ronglie and Xia (IJCAI-09) called this STV with parallel universes. ILLC Workshop on Collective Decision Making 2015 8 Stéphane Airiau (LAMSADE) - Voting in Parallel Universes

  9. Tree representation { a , b , c } b ⊲ c c ⊲ b { a , c } { a , b } b ⊲ a a ⊲ b { c } { a } { b } For each leaf nodes, we must count the number of tie-breaking rules that satisfy the “constraints” ➫ “counting the linear extensions” and it is a #-P complete problem. when ties are always between only two candidates, we can count in polynomial time. ILLC Workshop on Collective Decision Making 2015 9 Stéphane Airiau (LAMSADE) - Voting in Parallel Universes

  10. sketch of proof There is a tie between two candidates a and b at node V There are three types of constraints: x i ⊲ a (it cannot be a ⊲ x i as otherwise a would have been eliminated); let assume there are k such constraints y j ⊲ b ; let us assume l such constraints constraints that contains neither a nor b note that we cannot have x i = y j for all 1 � i � k , 1 � j � l we cannot have a constraint that include a x i and a y j (e.g. x i ⊲ y j or y j ⊲ x i for all 1 � i � k , 1 � j � l ) ➫ For the branch corresponding to the constraint a ⊲ b : count the number of sequences of length k + l + 2 for interspersing x 1 x 2 ... x k a with y 1 y 2 ... y l b such that a is before b . ➫ choose the position of k + 1 elements (corresponding to the x i and a ) among the k + l + 1 possible positions ➫ choose k + 1 elements from a set of k + l + 1 elements. ILLC Workshop on Collective Decision Making 2015 10 Stéphane Airiau (LAMSADE) - Voting in Parallel Universes

  11. case with ties between more than 2 candidates Assume a tie between three candidates, say a , b and c . say we eliminate a , the constraints down this branch are a ⊲ b and a ⊲ c . the following constraints are feasible: x i ⊲ a y j ⊲ b z k ⊲ c It is now possible to have a constraint x i y j as there could have been a tie between x i , y i and a in which x i is eliminated. our combinatorics argument will not work in this case. ILLC Workshop on Collective Decision Making 2015 11 Stéphane Airiau (LAMSADE) - Voting in Parallel Universes

  12. 3 candidates Anonynous Culture How often do we need to break a tie? y number of profiles number of anonymous profiles 6000 number profiles needing a tie-break number profiles with tie with 3 candidates 4000 2000 x 3 4 5 6 7 8 9 10 11 12 13 number of voters y 60 % of profiles 40 20 x 3 4 5 6 7 8 9 10 11 12 13 ILLC Workshop on Collective Decision Making 2015 12 Stéphane Airiau (LAMSADE) - Voting in Parallel Universes number of voters

  13. Sampling Impartial Culture y 100 90 80 70 % of profiles 60 50 6 candidates 40 30 5 candidates 20 4 candidates 10 3 candidates x 00 00 10 20 30 40 50 60 70 80 90 100 number of voters sampling 100,000 profiles with impartial culture number of times a tie-breaking rule is needed ILLC Workshop on Collective Decision Making 2015 13 Stéphane Airiau (LAMSADE) - Voting in Parallel Universes

  14. Decisiveness – Sampling Impartial Culture 3 candidates y 100 3 candidates argmax 90 3 candidates union percentage 80 70 60 50 40 x 30 0 5 10 15 20 25 number of voters 5 candidates y 100 5 candidates argmax 90 80 5 candidates union percentage 70 60 50 40 30 x 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 number of voters sampling 10,000 profiles with impartial there is a unique winner ILLC Workshop on Collective Decision Making 2015 14 Stéphane Airiau (LAMSADE) - Voting in Parallel Universes

  15. Future Work Banks: there is a polynomial algorithm to get one Banks winner ➫ union rule provides all Banks winner ➫ argmax rule discriminates among all Banks winner Complexity: we conjecture #-P complete for STV axiomatic: if the voting rule has some properties, what is conserved by union and argmax. Immediate for some axioms, not clear for others. Top Cycle tends to have a large winner set. Does the argmax rule helps to improve decisiveness? Comparison with perturbation method (Freeman, Conitzer, Brill AAMAS-15) Social Decision Scheme: we can propose particular SDS using our frequency rule. What are the properties of such rules? ILLC Workshop on Collective Decision Making 2015 15 Stéphane Airiau (LAMSADE) - Voting in Parallel Universes

Recommend


More recommend