Documente Academic
Documente Profesional
Documente Cultură
1/20
Introduction
2/20
Introduction
2/20
Weak priorities
3/20
Weak priorities
3/20
Weak priorities
3/20
Weak priorities
3/20
Weak priorities
3/20
An example of a weak priority for a school s:
Ps
Alice, Bob
Carol
Denis, Erin, Fred
Gilda
I Alice and Bob have a higher priority than any other student.
But Alice (Bob) doesn’t have a higher priority than Bob
(Alice).
I Carol has
I lower priority than Alice and Bob
I higher priority than Denis, Erin, Fred and Gilda.
4/20
An example of a weak priority for a school s:
Ps
Alice, Bob
Carol
Denis, Erin, Fred
Gilda
I Alice and Bob have a higher priority than any other student.
But Alice (Bob) doesn’t have a higher priority than Bob
(Alice).
I Carol has
I lower priority than Alice and Bob
I higher priority than Denis, Erin, Fred and Gilda.
4/20
An example of a weak priority for a school s:
Ps
Alice, Bob
Carol
Denis, Erin, Fred
Gilda
I Alice and Bob have a higher priority than any other student.
But Alice (Bob) doesn’t have a higher priority than Bob
(Alice).
I Carol has
I lower priority than Alice and Bob
I higher priority than Denis, Erin, Fred and Gilda.
4/20
To run an algorithm like DA we need strict priorities. Ties can be
(for instance) broken randomly.
5/20
To run an algorithm like DA we need strict priorities. Ties can be
(for instance) broken randomly.
5/20
To run an algorithm like DA we need strict priorities. Ties can be
(for instance) broken randomly.
5/20
To run an algorithm like DA we need strict priorities. Ties can be
(for instance) broken randomly.
5/20
To run an algorithm like DA we need strict priorities. Ties can be
(for instance) broken randomly.
5/20
Efficiency loss
So for s1 after breaking ties Bob has a strictly higher priority than
Carol
6/20
Efficiency loss
So for s1 after breaking ties Bob has a strictly higher priority than
Carol
6/20
Efficiency loss
So for s1 after breaking ties Bob has a strictly higher priority than
Carol
6/20
PAlice PBob PCarol πs1 πs2 πs3
s2 s3 s2 Alice Bob Carol
s1 s2 s3 Bob Alice Alice
s3 s1 s1 Carol Carol Bob
7/20
PAlice PBob PCarol
s2 s3 s2
s1 s2 s3
s3 s1 s1
9/20
We now have two sources of inefficiency:
I Due to the conflict between stability and Pareto efficiency (in
general they’re not compatible);
I Weak priorities can generate an assignment that is not the
student-optimal assignment.
9/20
We now have two sources of inefficiency:
I Due to the conflict between stability and Pareto efficiency (in
general they’re not compatible);
I Weak priorities can generate an assignment that is not the
student-optimal assignment.
9/20
We now have two sources of inefficiency:
I Due to the conflict between stability and Pareto efficiency (in
general they’re not compatible);
I Weak priorities can generate an assignment that is not the
student-optimal assignment.
9/20
We now have two sources of inefficiency:
I Due to the conflict between stability and Pareto efficiency (in
general they’re not compatible);
I Weak priorities can generate an assignment that is not the
student-optimal assignment.
9/20
We now have two sources of inefficiency:
I Due to the conflict between stability and Pareto efficiency (in
general they’re not compatible);
I Weak priorities can generate an assignment that is not the
student-optimal assignment.
9/20
Improvement cycles
10/20
Improvement cycles
10/20
Improvement cycles
10/20
Improvement cycles
10/20
Improvement cycles
10/20
Improvement cycles
10/20
Example
11/20
Example
PAlice PBob PCarol
s2 s3 s2
s1 s2 s3
s3 s1 s1
I Alice and Carol both want s2 . They have the same priority, so
they can point to Bob (enrolled at s2 ).
I Bob is the only one who want s3 , so he points to Carol.
11/20
Example
PAlice PBob PCarol
s2 s3 s2
s1 s2 s3
s3 s1 s1
I Alice and Carol both want s2 . They have the same priority, so
they can point to Bob (enrolled at s2 ).
I Bob is the only one who want s3 , so he points to Carol.
11/20
Example
PAlice PBob PCarol
s2 s3 s2
s1 s2 s3
s3 s1 s1
s2
Alice Bob
s3
s2
Carol
11/20
Example
Then we get
Using data from NYC, Abdulkadiroğlu, Pathak and Roth find that
for the years 2003–2007 they can improve on average the
assignment of about 1,700 students (around 2.5% of the students).
12/20
Theorem
If µ is a stable assignment and µ is Pareto dominated by another
assignment µ0 then there exists a stable improvement cycle.
Using data from NYC, Abdulkadiroğlu, Pathak and Roth find that
for the years 2003–2007 they can improve on average the
assignment of about 1,700 students (around 2.5% of the students).
12/20
DA + improvement cycles is not strategyproof.
Let
0
PAlice = s2 , s1 , s3
0
PBob = s2 , s1 , s3
13/20
DA + improvement cycles is not strategyproof.
Let
0
PAlice = s2 , s1 , s3
0
PBob = s2 , s1 , s3
13/20
DA + improvement cycles is not strategyproof.
Let
0
PAlice = s2 , s1 , s3
0
PBob = s2 , s1 , s3
13/20
For the profile
0
(PAlice , PBob , PCarol )
only µ is stable (not µ0 ).
14/20
For the profile
0
(PAlice , PBob , PCarol )
only µ is stable (not µ0 ).
14/20
For the profile
0
(PAlice , PBob , PCarol )
only µ is stable (not µ0 ).
14/20
For the profile
0
(PAlice , PBob , PCarol )
only µ is stable (not µ0 ).
14/20
For the profile
0
(PAlice , PBob , PCarol )
only µ is stable (not µ0 ).
14/20
How to break ties?
I Single tie-breaking:
The tie-breaking is the same for all schools.
Example: Alice and Bob are in the same tiers for schools s1
and s2 .
I If at s1 Alice ends up with a higher priority than Bob,
I then at s2 Alice ends up with a higher priority than Bob.
15/20
How to break ties?
I Single tie-breaking:
The tie-breaking is the same for all schools.
Example: Alice and Bob are in the same tiers for schools s1
and s2 .
I If at s1 Alice ends up with a higher priority than Bob,
I then at s2 Alice ends up with a higher priority than Bob.
15/20
How to break ties?
I Single tie-breaking:
The tie-breaking is the same for all schools.
Example: Alice and Bob are in the same tiers for schools s1
and s2 .
I If at s1 Alice ends up with a higher priority than Bob,
I then at s2 Alice ends up with a higher priority than Bob.
15/20
How to break ties?
I Single tie-breaking:
The tie-breaking is the same for all schools.
Example: Alice and Bob are in the same tiers for schools s1
and s2 .
I If at s1 Alice ends up with a higher priority than Bob,
I then at s2 Alice ends up with a higher priority than Bob.
15/20
How to break ties?
I Single tie-breaking:
The tie-breaking is the same for all schools.
Example: Alice and Bob are in the same tiers for schools s1
and s2 .
I If at s1 Alice ends up with a higher priority than Bob,
I then at s2 Alice ends up with a higher priority than Bob.
15/20
How to break ties?
I Single tie-breaking:
The tie-breaking is the same for all schools.
Example: Alice and Bob are in the same tiers for schools s1
and s2 .
I If at s1 Alice ends up with a higher priority than Bob,
I then at s2 Alice ends up with a higher priority than Bob.
15/20
Multiple tie-breaking seems more fair.
Then if Bob wants s he needs that Alice, Carol, Denis, Erin and
Fred are not assigned to s .
16/20
Multiple tie-breaking seems more fair.
Then if Bob wants s he needs that Alice, Carol, Denis, Erin and
Fred are not assigned to s .
16/20
Multiple tie-breaking seems more fair.
Then if Bob wants s he needs that Alice, Carol, Denis, Erin and
Fred are not assigned to s .
16/20
Multiple tie-breaking seems more fair.
Then if Bob wants s he needs that Alice, Carol, Denis, Erin and
Fred are not assigned to s .
16/20
So, multiple tie-breaking seems to give equal chances to each
student.
Theorem
If µ is a stable assignment such that:
I µ can be obtained using a multiple tie-breaking rule
I µ cannot be obtained using a single tie-breaking rule
then that assignment is not a student-optimal assignment.
17/20
So, multiple tie-breaking seems to give equal chances to each
student.
Theorem
If µ is a stable assignment such that:
I µ can be obtained using a multiple tie-breaking rule
I µ cannot be obtained using a single tie-breaking rule
then that assignment is not a student-optimal assignment.
17/20
Simulations using data from NYC
18/20
Constrained choice
19/20
Constrained choice
19/20
Constrained choice
19/20
Constrained choice
19/20
If choice is constrained, no assignment mechanism can be
strategyproof.
But if the unconstrained mechanism is strategyproof, we keep
something.
Theorem
If students cannot put more than k schools in their preferences and
the algorithm used is the DA or TTC:
20/20
If choice is constrained, no assignment mechanism can be
strategyproof.
But if the unconstrained mechanism is strategyproof, we keep
something.
Theorem
If students cannot put more than k schools in their preferences and
the algorithm used is the DA or TTC:
20/20
From very to less manipulable
21/20