bilevel problems, mpccs, and multi-leader-follower games · 2020. 10. 15. · didier aussel lab....
TRANSCRIPT
![Page 1: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/1.jpg)
Bilevel Problems, MPCCs, andMulti-Leader-Follower Games
Didier Aussel
Lab. Promes UPR CNRS 8521, University of Perpignan, France
ALOP autumn school - October 14th, 2020
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 2: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/2.jpg)
Professor in Applied Mathematics at Univ. of Perpignan
Perpignan, France
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 3: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/3.jpg)
Some personal data
Professor in Applied Mathematics at Univ. of Perpignan
Research topics:
- Bilevel programming, Nash games and in particularMulti-leader-follower games
- Energy management:
Electricity marketsIndustrial Eco-Parks (IEP)Demand-side management
- Variational and quasi-variational inequalities- Quasiconvex optimization
Research lab.: PROMES (CNRS)
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 4: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/4.jpg)
Some personal data
Professor in Applied Mathematics at Univ. of Perpignan
Research topics:
- Bilevel programming, Nash games and in particularMulti-leader-follower games
- Energy management:
Electricity marketsIndustrial Eco-Parks (IEP)Demand-side management
- Variational and quasi-variational inequalities- Quasiconvex optimization
Research lab.: PROMES (CNRS)
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 5: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/5.jpg)
Some personal data
Professor in Applied Mathematics at Univ. of Perpignan
Research topics:
- Bilevel programming, Nash games and in particularMulti-leader-follower games
- Energy management:
Electricity markets
Industrial Eco-Parks (IEP)Demand-side management
- Variational and quasi-variational inequalities- Quasiconvex optimization
Research lab.: PROMES (CNRS)
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 6: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/6.jpg)
Some personal data
Professor in Applied Mathematics at Univ. of Perpignan
Research topics:
- Bilevel programming, Nash games and in particularMulti-leader-follower games
- Energy management:
Electricity marketsIndustrial Eco-Parks (IEP)
Demand-side management
- Variational and quasi-variational inequalities- Quasiconvex optimization
Research lab.: PROMES (CNRS)
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 7: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/7.jpg)
Some personal data
Professor in Applied Mathematics at Univ. of Perpignan
Research topics:
- Bilevel programming, Nash games and in particularMulti-leader-follower games
- Energy management:
Electricity marketsIndustrial Eco-Parks (IEP)Demand-side management
- Variational and quasi-variational inequalities- Quasiconvex optimization
Research lab.: PROMES (CNRS)
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 8: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/8.jpg)
Some personal data
Professor in Applied Mathematics at Univ. of Perpignan
Research topics:
- Bilevel programming, Nash games and in particularMulti-leader-follower games
- Energy management:
Electricity marketsIndustrial Eco-Parks (IEP)Demand-side management
- Variational and quasi-variational inequalities
- Quasiconvex optimization
Research lab.: PROMES (CNRS)
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 9: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/9.jpg)
Some personal data
Professor in Applied Mathematics at Univ. of Perpignan
Research topics:
- Bilevel programming, Nash games and in particularMulti-leader-follower games
- Energy management:
Electricity marketsIndustrial Eco-Parks (IEP)Demand-side management
- Variational and quasi-variational inequalities- Quasiconvex optimization
Research lab.: PROMES (CNRS)
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 10: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/10.jpg)
Some personal data
Professor in Applied Mathematics at Univ. of Perpignan
Research topics:
- Bilevel programming, Nash games and in particularMulti-leader-follower games
- Energy management:
Electricity marketsIndustrial Eco-Parks (IEP)Demand-side management
- Variational and quasi-variational inequalities- Quasiconvex optimization
Research lab.: PROMES (CNRS)
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 11: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/11.jpg)
What do I work on?
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 12: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/12.jpg)
What do I work on?
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 13: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/13.jpg)
What do I work on?
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 14: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/14.jpg)
What do I work on?
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 15: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/15.jpg)
What do I work on?
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 16: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/16.jpg)
an advertisement
A short state of art on Multi-Leader-Follower games, D.A. and
A. Svensson, in a book dedicated to Stackelberg, editors A.
Zemkoho and S. Dempe, Springer Ed. (2019)
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 17: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/17.jpg)
Bilevel: some general comments
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 18: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/18.jpg)
BL: a �rst de�nition
A Bilevel Problem consists in an upper-level/leader'sproblem
�minx∈Rn� F (x, y)
s.t.
{x ∈ Xy ∈ S(x)
where ∅ 6= X ⊂ Rn and S(x) stands for the solution set of its
lower-level/follower's problem
miny∈Rm f(x, y)s.t g(x, y) ≤ 0
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 19: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/19.jpg)
A trivial example
Consider the following simple bilevel problem
�minx∈R� x
s.t.
{x ∈ [−1, 1]y ∈ S(x)
with S(x) = �y solving
miny∈R −xys.t x2(y2 − 1) ≤ 0 �
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 20: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/20.jpg)
A trivial example
Lower level problem:
miny∈R −x.ys.t x2(y2 − 1) ≤ 0 �
Note that the solution map of this convex problem is
S(x) :=
{1} x < 0{−1} x > 0R x = 0
Thus for each x 6= 0 there is a unique associated solution of the lower
level problem
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 21: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/21.jpg)
A trivial example
Lower level problem:
miny∈R −xys.t x2(y2 − 1) ≤ 0 �
Note that the solution map of this convex problem is
y
x
−∇F−1
1
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 22: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/22.jpg)
A trivial example
Consider the following simple bilevel problem
�minx∈R� −x.y
s.t.
{x ∈ [−1, 1]y ∈ S(x)
with S(x) = �y solving
S(x) :=
{1} x < 0{−1} x > 0R x = 0
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 23: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/23.jpg)
Ambiguity: Optimistic approach
An Optimistic Bilevel Problem consists in an
upper-level/leader's problem
minx∈Rn miny∈Rm F (x, y)
s.t.
{x ∈ Xy ∈ S(x)
where ∅ 6= X ⊂ Rn and S(x) stands for the solution set of its
lower-level/follower's problem
miny∈Rm f(x, y)s.t g(x, y) ≤ 0
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 24: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/24.jpg)
Ambiguity: Pessimistic approach
An Pessimistic Bilevel Problem consists in an
upper-level/leader's problem
minx∈Rn maxy∈Rm F (x, y)
s.t.
{x ∈ Xy ∈ S(x)
where ∅ 6= X ⊂ Rn and S(x) stands for the solution set of its
lower-level/follower's problem
miny∈Rm f(x, y)s.t g(x, y) ≤ 0
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 25: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/25.jpg)
Ambiguity: the most simple
And of course the "confortable situation" corresponds to the
case of a unique response
∀x ∈ X, S(x) = {y(x)}.
Then
minx∈Rn F (x, y(x))s.t.
{x ∈ X
For example when
for any x, g(x, ·) is quasiconvex and f(x, ·) is strictly convex.
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 26: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/26.jpg)
Ambiguity: the most simple
And of course the "confortable situation" corresponds to the
case of a unique response
∀x ∈ X, S(x) = {y(x)}.
Then
minx∈Rn F (x, y(x))s.t.
{x ∈ X
For example when
for any x, g(x, ·) is quasiconvex and f(x, ·) is strictly convex.
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 27: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/27.jpg)
Ambiguity: Selection approach
An "Selection-type" Bilevel Problem consists in an
upper-level/leader's problem
minx∈Rn F (x, y(x))
s.t.
{x ∈ Xy(x) is a uniquely determined selection of S(x)
J. Escobar & A. Jofré, Equilibrium Analysis of Electricity
Auctions (2011)
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 28: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/28.jpg)
Ambiguity: The new probabilistic approach
In one of the Elevator pitches (Monday), D.Salas and A.
Svensson proposed a probabilistic approach:
Consider a probability on the di�erent possible follower's
reactions
Minimize the expectation of the leader(s)
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 29: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/29.jpg)
An alternative point of view
Instead of considering the previous (optimistic) formulation of BL:
minx∈Rn miny∈Rm F (x, y)
s.t.
{x ∈ Xy ∈ S(x)
one can de�ne the (optimistic) value function
ϕmin(x) = miny{F (x, y) : g(x, y) ≤ 0} (1)
and the Bl problem becomes
minx∈Rn ϕmin(x)s.t. x ∈ X
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 30: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/30.jpg)
An alternative point of view
Instead of considering the previous (optimistic) formulation of BL:
minx∈Rn miny∈Rm F (x, y)
s.t.
{x ∈ Xy ∈ S(x)
one can de�ne the (optimistic) value function
ϕmin(x) = miny{F (x, y) : g(x, y) ≤ 0} (1)
and the Bl problem becomes
minx∈Rn ϕmin(x)s.t. x ∈ X
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 31: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/31.jpg)
An alternative point of view
Instead of considering the previous (pessimistic) formulation of BL:
minx∈Rn maxy∈Rm F (x, y)
s.t.
{x ∈ Xy ∈ S(x)
one can de�ne the (pessimistic) value function
ϕmax(x) = maxy{F (x, y) : g(x, y) ≤ 0} (2)
and the Bl problem becomes
minx∈Rn ϕmax(x)s.t. x ∈ X
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 32: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/32.jpg)
An alternative point of view
Instead of considering the previous (pessimistic) formulation of BL:
minx∈Rn maxy∈Rm F (x, y)
s.t.
{x ∈ Xy ∈ S(x)
one can de�ne the (pessimistic) value function
ϕmax(x) = maxy{F (x, y) : g(x, y) ≤ 0} (2)
and the Bl problem becomes
minx∈Rn ϕmax(x)s.t. x ∈ X
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 33: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/33.jpg)
An alternative point of view
This is the point of view presented in Stephan Dempe's book:
minx∈Rn min /maxy∈Rm F (x, y)
s.t.
{x ∈ Xy ∈ S(x)
vsminx∈Rn ϕmin/max(x)
s.t. x ∈ X
It immediately raises the question
What is a solution??
an optimal x = leader's optimal strategy?
an optimal couple (x, y) = couple of strategies of leader andfollower?
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 34: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/34.jpg)
An alternative point of view
This is the point of view presented in Stephan Dempe's book:
minx∈Rn min /maxy∈Rm F (x, y)
s.t.
{x ∈ Xy ∈ S(x)
vsminx∈Rn ϕmin/max(x)
s.t. x ∈ X
It immediately raises the question
What is a solution??
an optimal x = leader's optimal strategy?
an optimal couple (x, y) = couple of strategies of leader andfollower?
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 35: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/35.jpg)
An alternative point of view
This is the point of view presented in Stephan Dempe's book:
minx∈Rn min /maxy∈Rm F (x, y)
s.t.
{x ∈ Xy ∈ S(x)
vsminx∈Rn ϕmin/max(x)
s.t. x ∈ X
It immediately raises the question
What is a solution??
an optimal x = leader's optimal strategy?
an optimal couple (x, y) = couple of strategies of leader andfollower?
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 36: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/36.jpg)
Real life...
Actually usually when considering BL
minx∈Rn miny∈Rm F (x, y)
s.t.
{x ∈ Xy ∈ S(x)
people say
Step A: the leader plays �rst
Step B: the follower reacts
But in real life it's a little bit more complex....
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 37: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/37.jpg)
Real life...
Actually in real life, when considering BL
minx∈Rn miny∈Rm F (x, y)
s.t.
{x ∈ Xy ∈ S(x)
We only work for the leader!!
Indeed
the leader has a model of the follower's reaction: optimistic orpessimistic and
Step 1: we compute a solution x or (x, y) of the BL model
Step 2: the leader plays x
Step 3: the follower decides to play...whatever he wants!!!
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 38: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/38.jpg)
Real life...
Actually in real life, when considering BL
minx∈Rn miny∈Rm F (x, y)
s.t.
{x ∈ Xy ∈ S(x)
We only work for the leader!! Indeed
the leader has a model of the follower's reaction: optimistic orpessimistic and
Step 1: we compute a solution x or (x, y) of the BL model
Step 2: the leader plays x
Step 3: the follower decides to play...whatever he wants!!!
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 39: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/39.jpg)
An existence result (optimistic)
De�nition
The Mangasarian-Fromovitz constraint quali�cation (MFCQ) issatis�ed at (x, y) with y feasible point of the problem
miny{f(x, y) : g(x, y) ≤ 0}
if the system
∇ygi(x, y)d < 0 ∀ i ∈ I(x, y) := {j : gj(x, y) = 0}
has a solution.
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 40: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/40.jpg)
An existence result (cont.)
Assume that X = {x ∈ Rn : G(x) ≤ 0}
Theorem (Bank, Guddat, Klatte, Kummer, Tammer (83))
Let x with G(x) ≤ 0 be �xed.
the set {(x, y) : g(x, y) ≤ 0} is not empty and compact;
at each point (x, y) ∈ gphS with G(x) ≤ 0, assumption (MFCQ)is satis�ed;
then, the set-valued map S(·) is upper semicontinuous at (x, y) andthe function ϕo(·) is continuous at x.
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 41: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/41.jpg)
An existence result (cont.)
Theorem
Assume that
the set {(x, y) : g(x, y) ≤ 0} is not empty and compact;
at each point (x, y) ∈ gphS with G(x) ≤ 0, assumptions(MFCQ) is satis�ed;
the set {x : G(x) ≤ 0} is not empty and compact,
then optimistic bilevel problem has a (global) optimal solution.
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 42: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/42.jpg)
Bilevel problems and MPCC reformulation
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 43: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/43.jpg)
We consider a Bilevel Problem consisting in an upper-level /leader's problem
� minx∈Rn
� F (x, y)
s.t. y ∈ S(x), x ∈ X
where ∅ 6= X ⊂ Rn, and S(x) stands for the solution of its
lower-level / follower's problem
miny∈Rm
f(x, y)
s.t g(x, y) ≤ 0
which we assume to be convex and smooth, i.e. ∀x ∈ X, thefunctions f(x, ·) and gi(x, ·) are smooth convex functions, and
the gradients ∇ygi,∇yf are continuous, i = 1, ..., p.
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 44: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/44.jpg)
MPCC reformulation
Replacing the lower-level problem by its KKT conditions, gives placeto a Mathematical Program with Complementarity Constraints.
Bilevel
�minx∈X
�F (x, y)
s.t. y ∈ S(x)
with S(x) = �y solving
miny∈Rm
f(x, y)
s.t g(x, y) ≤ 0 �
MPCC
�minx∈X
�F (x, y)
s.t. (y, u) ∈ KKT (x)
with KKT (x) = �(y, u) solving
{∇yf(x, y) + uT∇yg(x, y) = 0
0 ≤ u ⊥ −g(x, y) ≥ 0 �
We write Λ(x, y) for the set of u satisfying (y, u) ∈ KKT (x).
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 45: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/45.jpg)
Example 1
Consider the following Bilevel problem and its MPCC reformulation
Bilevel
� minx∈[−1,1]
� x
s.t. y ∈ S(x)
with S(x) = �y solving
miny∈R
xy
s.t x2(y2 − 1) ≤ 0 �
MPCC
� minx∈[−1,1]
� x
s.t. (y, u) ∈ KKT (x)
with KKT (x) = �(y, u) solving
{x + u · 2yx2 = 00 ≤ u ⊥ −x2(y2 − 1) ≥ 0 �
1 (0,−1, u) is a local solution of �MPCC�, for any u ∈ Λ(0,−1) = R+
2 (0,−1) is NOT a local solution of �Bilevel�
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 46: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/46.jpg)
x
y
uKKT (·)
−∇F
(a) (0,−1, u) is alocal solution ofMPCC, ∀u ∈ R+.
y
x
−∇F−1
1
S(·)
(b) (0,−1) isn't alocal solution of theBilevel problem.
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 47: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/47.jpg)
Optimistic and Pessimistic approaches
The optimistic Bilevel (OB) is
minx
miny
F (x, y)
s.t. y ∈ S(x), x ∈ X.
The pessimistic Bilevel (PB) is
minx
maxy
F (x, y)
s.t. y ∈ S(x), x ∈ X.
The optimistic MPCC (OMPCC):
minx
miny
F (x, y)
s.t. (y, u) ∈ KKT (x), x ∈ X.
The pessimistic MPCC (PMPCC):
minx
maxy
F (x, y)
s.t. (y, u) ∈ KKT (x), x ∈ X.
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 48: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/48.jpg)
Optimistic and Pessimistic approaches
The optimistic Bilevel (OB) is
minx
miny
F (x, y)
s.t. y ∈ S(x), x ∈ X.
The pessimistic Bilevel (PB) is
minx
maxy
F (x, y)
s.t. y ∈ S(x), x ∈ X.
The optimistic MPCC (OMPCC):
minx
miny
F (x, y)
s.t. (y, u) ∈ KKT (x), x ∈ X.
The pessimistic MPCC (PMPCC):
minx
maxy
F (x, y)
s.t. (y, u) ∈ KKT (x), x ∈ X.
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 49: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/49.jpg)
Optimistic approachIs bilevel programming a special case of a MPCC?
S. Dempe -J. Dutta (2012 Math. Prog.)
minx
minyF (x, y)
s.t. y ∈ S(x), x ∈ X.
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 50: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/50.jpg)
Local solutions for in optimistic approach
De�nition
A local (resp. global) solution of (OB) is a point
(x, y) ∈ Gr(S) if there exists U ∈ N (x, y) (resp. U = Rn × Rm)such that
F (x, y) ≤ F (x, y), ∀(x, y) ∈ U ∩Gr(S).
De�nition
A local (resp. global) solution for (OMPCC) is a triplet
(x, y, u) ∈ Gr(KKT ) such that there exists U ∈ N (x, y, u)(resp. U = Rn × Rm × Rp) with
F (x, y) ≤ F (x, y), ∀(x, y, u) ∈ U ∩Gr(KKT ).
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 51: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/51.jpg)
Results for the optimistic case
In Dempe-Dutta it was considered the Slater type constraint
quali�cation for a parameter x ∈ X:
Slater: ∃y(x) ∈ Rm s.t. gi(x, y(x)) < 0, ∀i = 1, .., p.
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 52: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/52.jpg)
Results for the optimistic case
Theorem 1 Dempe-Dutta (2012)
Assume the convexity condition and Slater's CQ at x.
1 If (x, y) is a local solution for (OB), then for each
u ∈ Λ(x, y), (x, y, u) is a local solution for (OMPCC).
2 Conversely, assume that Slater's CQ holds on a
neighbourhood of x, Λ(x, y) 6= ∅, and (x, y, u) is a local
solution of (OMPCC) for every u ∈ Λ(x, y). Then (x, y) is alocal solution of (OB).
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 53: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/53.jpg)
Under the convexity assumption and some CQ ensuring KKT (x) 6= ∅,∀x ∈ X:
(x, y, u) sol
of (OMPCC)
(x, y) sol
of (OB)
∀u ∈ Λ(x, y)
Figure: Global solution comparison in optimistic approach
(x, y) local
sol of (OB)
∀u ∈ Λ(x, y),
(x, y, u)
local sol of
(OMPCC)if Slater's CQ
holds around x
Figure: Local solution comparison in optimistic approach
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 54: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/54.jpg)
Example 1 (optimistic)
Consider the following optimistic Bilevel problem
minx∈[−1,1]
minyx
s.t. y ∈ S(x), x ∈ R
with lower-level
miny− xy
s.t x2(y2 − 1) ≤ 0.
1 (0,−1, u) is a local solution of (OMPCC), for any
u ∈ Λ(0,−1) = R+
2 (0,−1) is NOT a local solution of (OB).
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 55: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/55.jpg)
Pessimistic ApproachIs bilevel programming a special case of a (MPCC)?
Aussel - Svensson (2019 - J. Optim. Theory Appl.)
minx
maxy
F (x, y)
s.t. y ∈ S(x), x ∈ X.
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 56: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/56.jpg)
De�nition
A pair (x, y) is said to be a local (resp. global) solution for (PB),
if (x, y) ∈ Gr(Sp) and ∃U ∈ N (x, y) such that
F (x, y) ≤ F (x, y), ∀(x, y) ∈ U ∩Gr(Sp). (3)
where Sp(x) := argmaxy {F (x, y) | y ∈ S(x)} .
De�nition
A triplet (x, y, u) is said to be a local (resp. global) solution for
(PMPCC), if (x, y, u) ∈ Gr(KKTp) and ∃U ∈ N (x, y, u) suchthat
F (x, y) ≤ F (x, y), ∀(x, y, u) ∈ U ∩Gr(KKTp). (4)
where KKTp(x) := argmaxy,u {F (x, y) | (y, u) ∈ KKT (x)} .
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 57: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/57.jpg)
Results for the pessimistic case
Theorem 2
Assume the convexity condition and that KKT (x) 6= ∅, ∀x ∈ X.
1 If (x, y) is a local solution for (PB), then for each
u ∈ Λ(x, y), (x, y, u) is a local solution for (PMPCC).2 Conversely, assume that one of the following condition are
satis�ed:
1 The multifunction KKTp is LSC around (x, y, u) and(x, y, u) is a local solution of (PB).
2 Slater's CQ holds on a neighbourhood of x, Λ(x, y) 6= ∅, andfor every u ∈ Λ(x, y), (x, y, u) is a local solution of(PMPCC).
Then (x, y) is a local solution of (PB).
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 58: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/58.jpg)
Example 1 (pessimistic)
Consider the following pessimistic Bilevel problem
minx∈[−1,1]
maxy
x
s.t. y ∈ S(x), x ∈ R
with lower-level
miny− xy
s.t x2(y2 − 1) ≤ 0.
1 (0,−1, u) is a local solution of (PMPCC), for any
u ∈ Λ(0,−1) = R+
2 (0,−1) is NOT a local solution of (PB).
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 59: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/59.jpg)
Example 2
Consider the following Bilevel problem
� minx
�x
s.t. y ∈ S(x)
with S(x) the solution of the lower-level problem
miny{−y | x+ y ≤ 0, y ≤ 0}
Even though Slater's CQ holds, we have
1 (0, 0, u1, u2) with (u1, u2) ∈ Λ(0, 0) = {(λ, 1− λ) | λ ∈ [0, 1]} is alocal solution of �(MPCC)�, i� u1 6= 0,
2 (0, 0) is NOT a local solution for �(B)�.
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 60: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/60.jpg)
Under the convexity assumption and some (CQ) ensuring KKT (x) 6= ∅,∀x ∈ X:
(x, y, u) sol
of (PMPCC)
(x, y) sol
of (PB)
∀u ∈ Λ(x, y)
Figure: Global solution comparison in pessimistic approach
(x, y) local
sol of (PB)
∀u ∈ Λ(x, y),
(x, y, u)
local sol of
(PMPCC)Slater's CQ for
all x around x
Figure: Local solutions comparison in pessimistic approach
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 61: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/61.jpg)
An introduction to MLFG
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 62: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/62.jpg)
Notations and de�nitions: Nash problems
A Nash equilibrium problem is a noncooperative game inwhich the decision function (cost/bene�t) of each playerdepends on the decision of the other players.
Denote by N the number of players and each player i controlsvariables xi ∈ Rni . The �total strategy vector� is x which will be oftendenoted by
x = (xi, x−i).
where x−i is the strategy vector of the other players.
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 63: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/63.jpg)
Notations and de�nitions: Nash problems
A Nash equilibrium problem is a noncooperative game inwhich the decision function (cost/bene�t) of each playerdepends on the decision of the other players.
Denote by N the number of players and each player i controlsvariables xi ∈ Rni . The �total strategy vector� is x which will be oftendenoted by
x = (xi, x−i).
where x−i is the strategy vector of the other players.
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 64: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/64.jpg)
Nash Equilibrium Problem (NEP)
The strategy of player i belongs to a strategy set
xi ∈ Xi
Given the strategies x−i of the other players, the aim of player iis to choose a strategy xi solving
Pi(x−i) max θi(x
i, x−i)
s.t. xi ∈ Xi
where θi(·, x−i) : Rni → R is the decision function for player i.
A vector x is a Nash Equilibrium if
for any i, xi solves Pi(x−i).
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 65: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/65.jpg)
Nash Equilibrium Problem (NEP)
The strategy of player i belongs to a strategy set
xi ∈ Xi
Given the strategies x−i of the other players, the aim of player iis to choose a strategy xi solving
Pi(x−i) max θi(x
i, x−i)
s.t. xi ∈ Xi
where θi(·, x−i) : Rni → R is the decision function for player i.
A vector x is a Nash Equilibrium if
for any i, xi solves Pi(x−i).
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 66: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/66.jpg)
Nash Equilibrium Problem (NEP)
The strategy of player i belongs to a strategy set
xi ∈ Xi
Given the strategies x−i of the other players, the aim of player iis to choose a strategy xi solving
Pi(x−i) max θi(x
i, x−i)
s.t. xi ∈ Xi
where θi(·, x−i) : Rni → R is the decision function for player i.
A vector x is a Nash Equilibrium if
for any i, xi solves Pi(x−i).
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 67: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/67.jpg)
Generalized Nash Equilibrium Problem
A generalized Nash equilibrium problem (GNEP) is anoncooperative game in which the decision function and
strategy set of each player depend on the decision of the otherplayers.
The strategy of player i belongs to a strategy set
xi ∈ Xi(x
−i)
which depends on the decision variables of the other players.
Given the strategies x−i of the other players, the aim of player i is to choose a strategy
xi solving
Pi(x−i) max θi(x
i, x−i)
s.t. xi ∈ Xi(x−i)
where θi(·, x−i) : Rni → R is the decision function for player i.
A vector x is a Generalized Nash Equilibrium if
for any i, xi
solves Pi(x−i
).
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 68: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/68.jpg)
Generalized Nash Equilibrium Problem
A generalized Nash equilibrium problem (GNEP) is anoncooperative game in which the decision function and
strategy set of each player depend on the decision of the otherplayers.
The strategy of player i belongs to a strategy set
xi ∈ Xi(x
−i)
which depends on the decision variables of the other players.
Given the strategies x−i of the other players, the aim of player i is to choose a strategy
xi solving
Pi(x−i) max θi(x
i, x−i)
s.t. xi ∈ Xi(x−i)
where θi(·, x−i) : Rni → R is the decision function for player i.
A vector x is a Generalized Nash Equilibrium if
for any i, xi
solves Pi(x−i
).
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 69: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/69.jpg)
Generalized Nash Equilibrium Problem
A generalized Nash equilibrium problem (GNEP) is anoncooperative game in which the decision function and
strategy set of each player depend on the decision of the otherplayers.
The strategy of player i belongs to a strategy set
xi ∈ Xi(x
−i)
which depends on the decision variables of the other players.
Given the strategies x−i of the other players, the aim of player i is to choose a strategy
xi solving
Pi(x−i) max θi(x
i, x−i)
s.t. xi ∈ Xi(x−i)
where θi(·, x−i) : Rni → R is the decision function for player i.
A vector x is a Generalized Nash Equilibrium if
for any i, xi
solves Pi(x−i
).
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 70: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/70.jpg)
General model
Generalized Nash game (GNEP):
minx1
θ1(x)
s.t.{x1 ∈ X1(x−1)
. . .min
xnθn(x)
s.t.{xn ∈ Xn(x−n)
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 71: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/71.jpg)
A classical existence result
Theorem (Ichiishi-Quinzii 1983)
Let a GNEP be given and suppose that
1 For each ν = 1, ..., N there exist a nonempty, convex and
compact set Kν ⊂ Rnν such that the point-to-set map
Xν : K−ν ⇒ Kν , is both upper and lower semicontinuous
with nonempty closed and convex values, where
K−ν :=∏ν′ 6=ν Kν .
2 For every player ν, the function θν is continuous and
θν(·, x−ν) is quasi-convex on Xν(x−ν).
Then a generalized Nash equilibrium exists.
Note that in Aussel-Dutta (2008) an alternative proof of
existence of equilibria has been given, under the assumption of
the Rosen's law, by using the normal approach technique.
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 72: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/72.jpg)
Structure of the set of GNEPs
Example
Let x = (x1, x2) ∈ [0, 4]2 and fν(x) := dTν (x)2, where T1 is thetriangle with vertices (0, 0), (0, 4) and (1, 2), and T2 is the trianglewhose vertices are (0, 0), (4, 0) and (2, 1). LetSν(x−ν) := argminxν
{fν(x1, x2) | xν ∈ [0, 4]
}. We see that
S1(x2) ={x1 ∈ [0, 4] | (x1, x2) ∈ T1
}for x2 ∈ [0, 1]
S1(x2) = {2} for all x2 ∈ (1, 4])
S2(x1) ={x2 ∈ [0, 4] | (x1, x2) ∈ T2
}for x1 ∈ [0, 1]
S2(x1) = {2} for all x1 ∈ (1, 4]).
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 73: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/73.jpg)
Structure of the set of GNEPs (cont.)
x1
x2
S1(·)
S2(·)
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 74: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/74.jpg)
A particular case
Multi-Leader-Follower-Game (MLFG):
minx1y1,..,yp
θ1(x, y)
s.t.
{x1 ∈ X1(x−1)y ∈ Y (x)
. . .
minxny1,..,yp
θn(x, y)
s.t.
{xn ∈ Xn(x−n)y ∈ Y (x)
↓↑ ↓↑
miny1,..,yp φ1(x, y)s.t.
{y ∈ Y (x)
. . .miny1,..,yp φp(x, y)
s.t.{y ∈ Y (x)
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 75: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/75.jpg)
and another problem
Single-Leader-Multi-Follower-Game (SLMFG):
minxy1,..,yp
θ1(x, y)
s.t.
{x ∈ Xy ∈ Y (x)
↓↑
miny1,..,yp φ1(x, y)
s.t.{y ∈ Y (x)
. . .miny1,..,yp φp(x, y)
s.t.{y ∈ Y (x)
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 76: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/76.jpg)
A particular case
Multi-Leader-Single-Follower-Game (MLSFG):
minx1y1,..,yp
θ1(x, y)
s.t.
{x1 ∈ X1(x−1)y ∈ Y (x)
. . .
minxny1,..,yp
θn(x, y)
s.t.
{xn ∈ Xn(x−n)y ∈ Y (x)
↓↑
miny1,..,yp φ1(x, y)s.t.
{y ∈ Y (x)
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 77: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/77.jpg)
MLSF game: ill-posedness
minx1,y θ1(x1, x2, y) = x1.y
s.t.
{x1 ∈ [0, 1]y ∈ S(x1, x2)
minx2,y θ1(x1, x2, y) = −x2.y
s.t.
{x2 ∈ [0, 1]y ∈ S(x1, x2)
with
miny f(x1, x2, y) = 13y
3 − (x1 + x2)2y
s.t. y ∈ R
Exercise: Please analyse this small example...
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 78: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/78.jpg)
MLSF game: ill-posedness
minx1,y θ1(x1, x2, y) = x1.y
s.t.
{x1 ∈ [0, 1]y ∈ S(x1, x2)
minx2,y θ1(x1, x2, y) = −x2.y
s.t.
{x2 ∈ [0, 1]y ∈ S(x1, x2)
with
miny f(x1, x2, y) = 13y
3 − (x1 + x2)2y
s.t. y ∈ R
Exercise: Please analyse this small example...
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 79: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/79.jpg)
MLSF game: ill-posedness
The follower problem �rst
miny f(x1, x2, y) = 13y
3 − (x1 + x2)2y
s.t. y ∈ R
The solution map of this follower problem is
S(x1, x2) = {y1 = x1 + x2, y2 = −x1 − x2}.
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 80: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/80.jpg)
MLSF game: ill-posedness
The follower problem �rst
miny f(x1, x2, y) = 13y
3 − (x1 + x2)2y
s.t. y ∈ R
The solution map of this follower problem is
S(x1, x2) = {y1 = x1 + x2, y2 = −x1 − x2}.
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 81: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/81.jpg)
MLSF game: ill-posedness
The solution map of this follower problem is
S(x1, x2) = {y1 = x1 + x2, y2 = −x1 − x2}.
The leader 1 problem
θ1(x, y) = x1.y =
{x21 + x1.x2 if y = y1−x21 − x1.x2 if y = y2
Thus the response function of player 1 is
R1(x2) =
{{0} if y = y1 with a payo� = 0{1} if y = y2 with a payo� = −1− x2
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 82: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/82.jpg)
MLSF game: ill-posedness
The solution map of this follower problem is
S(x1, x2) = {y1 = x1 + x2, y2 = −x1 − x2}.
The leader 1 problem
θ1(x, y) = x1.y =
{x21 + x1.x2 if y = y1−x21 − x1.x2 if y = y2
Thus the response function of player 1 is
R1(x2) =
{{0} if y = y1 with a payo� = 0{1} if y = y2 with a payo� = −1− x2
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 83: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/83.jpg)
MLSF game: ill-posedness
The solution map of this follower problem is
S(x1, x2) = {y1 = x1 + x2, y2 = −x1 − x2}.
The leader 2 problem
θ1(x, y) = −x2.y =
{−x21 − x1.x2 if y = y1x21 + x1.x2 if y = y2
Thus the response function of player 1 is
R2(x1) =
{{1} if y = y1 with a payo� = −1− x1{0} if y = y2 with a payo� = 0
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 84: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/84.jpg)
MLSF game: ill-posedness
The solution map of this follower problem is
S(x1, x2) = {y1 = x1 + x2, y2 = −x1 − x2}.
The leader 2 problem
θ1(x, y) = −x2.y =
{−x21 − x1.x2 if y = y1x21 + x1.x2 if y = y2
Thus the response function of player 1 is
R2(x1) =
{{1} if y = y1 with a payo� = −1− x1{0} if y = y2 with a payo� = 0
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 85: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/85.jpg)
MLSF game: ill-posedness
R1(x2) =
{{(0, y = y1)} with a payo� = 0{(1, y = y2)} with a payo� = −1− x2
R2(x1) =
{{(1, y = y1)} with a payo� = −1− x1{(0, y = y2)} with a payo� = 0
So the Nash equilibrium will be (x1, x2) = (1, 1) but....
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 86: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/86.jpg)
MLSF game: ill-posedness
R1(x2) =
{{(0, y = y1)} with a payo� = 0{(1, y = y2)} with a payo� = −1− x2
R2(x1) =
{{(1, y = y1)} with a payo� = −1− x1{(0, y = y2)} with a payo� = 0
So the Nash equilibrium will be (x1, x2) = (1, 1) but....
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 87: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/87.jpg)
A �nal model
For the Demand-side management, we recently introduced theMulti-Leader-Disjoint-Follower game
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 88: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/88.jpg)
Just one example [Pang-Fukushima 05]
Let us consider a 2-leader-single-follower game:
minx1,y12x1 + y minx2,y − 1
2x2 − y{
x1 ∈ [0, 1]y ∈ S(x1, x2)
{x2 ∈ [0, 1]y ∈ S(x1, x2)
where S(x1, x2) is the solution map of
miny≥0 y(−1 + x1 + x2) +1
2y2
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 89: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/89.jpg)
Just one example [Pang-Fukushima 05]
Let us consider a 2-leader-single-follower game:
minx1,y112x1 + y1 minx2,y2 − 1
2x2 − y2{
x1 ∈ [0, 1]y1 ∈ S(x1, x2)
{x2 ∈ [0, 1]y2 ∈ S(x1, x2)
where S(x1, x2) is the solution map of
miny≥0 y(−1 + x1 + x2) +1
2y2
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 90: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/90.jpg)
Actually S(x1, x2) = max{0, 1− x1 − x2} thus the problem becomes
minx1,y112x1 + y1 minx2,y2 − 1
2x2 − y2{
x1 ∈ [0, 1]y1 = max{0, 1− x1 − x2}
{x2 ∈ [0, 1]y2 = max{0, 1− x1 − x2}
Then the Response maps are
R1(x2) = {1− x2} and R2(x1) =
{0} x1 ∈ [0, 1
2[
{0, 1} x1 = 12
{1} x1 ∈] 12, 1]
and thus there is no Nash equilibrium.......
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 91: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/91.jpg)
Actually S(x1, x2) = max{0, 1− x1 − x2} thus the problem becomes
minx1,y112x1 + y1 minx2,y2 − 1
2x2 − y2{
x1 ∈ [0, 1]y1 = max{0, 1− x1 − x2}
{x2 ∈ [0, 1]y2 = max{0, 1− x1 − x2}
Then the Response maps are
R1(x2) = {1− x2} and R2(x1) =
{0} x1 ∈ [0, 1
2[
{0, 1} x1 = 12
{1} x1 ∈] 12, 1]
and thus there is no Nash equilibrium.......
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 92: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/92.jpg)
But let us consider the slightly modi�ed problem.......
minx1,y112x1 + y1 minx2,y2 − 1
2x2 − y2 x1 ∈ [0, 1]
y1 = max{0, 1− x1 − x2}y2 = max{0, 1− x1 − x2}
x2 ∈ [0, 1]y1 = max{0, 1− x1 − x2}y2 = max{0, 1− x1 − x2}
that can be proved to have a (unique) Nash equilibrium namely
(x1, x2) = (0, 1) with y1 = y2 = 0!!!!
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 93: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/93.jpg)
But let us consider the slightly modi�ed problem.......
minx1,y112x1 + y1 minx2,y2 − 1
2x2 − y2 x1 ∈ [0, 1]
y1 = max{0, 1− x1 − x2}y2 = max{0, 1− x1 − x2}
x2 ∈ [0, 1]y1 = max{0, 1− x1 − x2}y2 = max{0, 1− x1 − x2}
that can be proved to have a (unique) Nash equilibrium namely
(x1, x2) = (0, 1) with y1 = y2 = 0!!!!
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 94: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/94.jpg)
The kind of �trick� is called �All Equilibrium approach� and has beenintroduced in A.A. Kulkarni & U.V. Shanbhag, A Shared-ConstraintApproach to Multi-Leader Multi-Follower Games, Set-Valued Var. Anal(2014).
They proved that every Nash equilibirum (initial problem) is a Nashequilibrium for the �all equilibrium� formulation.
It corresponds to the case where each leader takes into account the
conjectures regarding the follower decision made by all other leaders....
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 95: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/95.jpg)
Some motivation examples
Electricity markets
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 96: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/96.jpg)
A short introduction to electricity markets
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 97: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/97.jpg)
A short introduction to electricity markets (cont.)
Volume of exchanges
Bid schedule of the spot market
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 98: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/98.jpg)
A short introduction to electricity markets (cont.)
Volume of exchanges
Bid schedule of the spot market
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 99: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/99.jpg)
Modeling an Electricity Markets
electricity market consists of
i) generators/consumers i ∈ N respect their own interests incompetition with others
ii) market operator (ISO) who maintain energy generation andload balance, and protect public welfare
the ISO has to consider:
ii) quantities qi of generated/consumed electricityiii) electricity dispatch te with respect to transmission
capacities
since 1990s, Generalized Nash equilibrium problem is the mostpopular way of modeling spot electricity markets or, more precisely,Multi-leader-common-follower game
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 100: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/100.jpg)
Modeling an Electricity Markets
electricity market consists of
i) generators/consumers i ∈ N respect their own interests incompetition with others
ii) market operator (ISO) who maintain energy generation andload balance, and protect public welfare
the ISO has to consider:
ii) quantities qi of generated/consumed electricityiii) electricity dispatch te with respect to transmission
capacities
since 1990s, Generalized Nash equilibrium problem is the mostpopular way of modeling spot electricity markets or, more precisely,Multi-leader-common-follower game
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 101: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/101.jpg)
Multi-Leader-Common-Follower game
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 102: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/102.jpg)
Multi-Leader-Common-Follower game
A classical problem (of a producer) is the best response search
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 103: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/103.jpg)
Models with linear unit bid functions
Electricity markets without transmission losses:
X. Hu & D. Ralph, Using EPECs to Model Bilevel Games inRestructured Electricity Markets with Locational Prices, OperationsResearch (2007). bid-on-a-only
Electricity markets with transmission losses:
Henrion, R., Outrata, J. & Surowiec, T., Analysis ofM-stationary points to an EPEC modeling oligopolisticcompetition in an electricity spot market, ESAIM: COCV(2012). M-stationary pointsD. A., R. Correa & M. Marechal Spot electricity marketwith transmission losses, J. Industrial Manag. Optim(2013). existence of Nash equil., case of a two island modelD.A., M. Cervinka & M. Marechal, Deregulated electricitymarkets with thermal losses and production bounds,RAIRO (2016) production bounds, well-posedness of model
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 104: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/104.jpg)
Models with linear unit bid functions
Electricity markets without transmission losses:
X. Hu & D. Ralph, Using EPECs to Model Bilevel Games inRestructured Electricity Markets with Locational Prices, OperationsResearch (2007). bid-on-a-only
Electricity markets with transmission losses:
Henrion, R., Outrata, J. & Surowiec, T., Analysis ofM-stationary points to an EPEC modeling oligopolisticcompetition in an electricity spot market, ESAIM: COCV(2012). M-stationary pointsD. A., R. Correa & M. Marechal Spot electricity marketwith transmission losses, J. Industrial Manag. Optim(2013). existence of Nash equil., case of a two island modelD.A., M. Cervinka & M. Marechal, Deregulated electricitymarkets with thermal losses and production bounds,RAIRO (2016) production bounds, well-posedness of model
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 105: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/105.jpg)
Some references on the topic (cont.)
Best response in electricity markets:
E. Anderson and A. Philpott, Optimal O�er Constructionin Electricity Markets, Mathematics of Operations Research(2002). Linear bid function - necessary optimality cond. forlocal best response in time dependent caseD. Aussel, P. Bendotti and M. Pi²t¥k, Nash Equilibrium inPay-as-bid Electricity Market : Part 2 - Best Response ofProducer, Optimization (2017) linear unit bid function,explicit formula for best response
Explicit formula for equilibria
D. Aussel, P. Bendotti and M. Pi²t¥k, Nash Equilibrium in Pay-as-bidElectricity Market : Part 1 - Existence and Characterisation,Optimization (2017) explicit formula for equilibria
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 106: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/106.jpg)
Some references on the topic (cont.)
Best response in electricity markets:
E. Anderson and A. Philpott, Optimal O�er Constructionin Electricity Markets, Mathematics of Operations Research(2002). Linear bid function - necessary optimality cond. forlocal best response in time dependent caseD. Aussel, P. Bendotti and M. Pi²t¥k, Nash Equilibrium inPay-as-bid Electricity Market : Part 2 - Best Response ofProducer, Optimization (2017) linear unit bid function,explicit formula for best response
Explicit formula for equilibria
D. Aussel, P. Bendotti and M. Pi²t¥k, Nash Equilibrium in Pay-as-bidElectricity Market : Part 1 - Existence and Characterisation,Optimization (2017) explicit formula for equilibria
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 107: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/107.jpg)
But also...
Non a priori structured bid functions
Escobar, J.F. and Jofré, A., Monopolistic competition inelectricity networks with resistance losses, Econom. Theory44 (2010).Escobar, J.F. and Jofré, A., Equilibrium analysis ofelectricity auctions, preprint (2014).E. Anderson, P. Holmberg and A. Philpott, Mixed strategiesin discriminatory divisible-good auctions, The RANDJournal of Economics (2013). necessary optimality cond.for local best response
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 108: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/108.jpg)
Notations
D. Aussel, P. Bendotti and M. Pi²t¥k, Nash Equilibrium in Pay-as-bid Electricity Market :Optimization
Part 1 - Existence and Characterisation, 66:6 (2017)
Part 2 - Best Response of Producer, 66:6 (2017).
Let consider a �xed time instant and denote
D > 0 be the overall energy demand of all consumers
N be the set of producers
qi ≥ 0 be the production of i-th producer, i ∈ N
We assume that producer i ∈ N provides to the ISO a quadratic bidfunction aiqi + biq
2i given by ai, bi ≥ 0.
Similarly, let Aiqi +Biq2i be the true production cost of i-th producer
with Ai ≥ 0 and Bi > 0 re�ecting the increasing marginal cost of
production.
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 109: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/109.jpg)
Notations
D. Aussel, P. Bendotti and M. Pi²t¥k, Nash Equilibrium in Pay-as-bid Electricity Market :Optimization
Part 1 - Existence and Characterisation, 66:6 (2017)
Part 2 - Best Response of Producer, 66:6 (2017).
Let consider a �xed time instant and denote
D > 0 be the overall energy demand of all consumers
N be the set of producers
qi ≥ 0 be the production of i-th producer, i ∈ N
We assume that producer i ∈ N provides to the ISO a quadratic bidfunction aiqi + biq
2i given by ai, bi ≥ 0.
Similarly, let Aiqi +Biq2i be the true production cost of i-th producer
with Ai ≥ 0 and Bi > 0 re�ecting the increasing marginal cost of
production.
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 110: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/110.jpg)
Notations
D. Aussel, P. Bendotti and M. Pi²t¥k, Nash Equilibrium in Pay-as-bid Electricity Market :Optimization
Part 1 - Existence and Characterisation, 66:6 (2017)
Part 2 - Best Response of Producer, 66:6 (2017).
Let consider a �xed time instant and denote
D > 0 be the overall energy demand of all consumers
N be the set of producers
qi ≥ 0 be the production of i-th producer, i ∈ N
We assume that producer i ∈ N provides to the ISO a quadratic bidfunction aiqi + biq
2i given by ai, bi ≥ 0.
Similarly, let Aiqi +Biq2i be the true production cost of i-th producer
with Ai ≥ 0 and Bi > 0 re�ecting the increasing marginal cost of
production.
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 111: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/111.jpg)
Multi-Leader-Common-Follower game
Peculiarity of electricity markets is their bi-level structure:
Pi(a−i, b−i, D) maxai,bi
maxqi
aiqi + biq2i − (Aiqi + Biq
2i )
such that
{ai, bi ≥ 0
(qj)j∈N ∈ Q(a, b)
where set-valued mapping Q(a, b) denotes solution set of
ISO(a, b,D) Q(a, b) = argminq
∑i∈N (aiqi + biq
2i )
such that
qi ≥ 0 , ∀i ∈ N∑i∈N
qi = D
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 112: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/112.jpg)
Some motivation examples
Industrial Eco-Parks
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 113: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/113.jpg)
What is an � Eco-park � ?
Example of water management
In a geographical area,there are di�erentcompanies 1, . . . , n
Each of them is buyingfresh water (high price)for their productionprocesses
Each company generatessome "dirty water" andhave to pay for discharge
Stand alone situation
Company 2
Company 3
Fresh water
Discharge water
Company 1Company 1
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 114: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/114.jpg)
How dos it work ?
The aims in designing Industrial Eco-park (IEP) are
a) Reduce cost of production of each company
b) Reduce the environmental impact of the whole production
Thus "Eco" of IEP is at the same time Economical and ecological
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 115: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/115.jpg)
What is an � Eco-park � ?
Example of water management
How to reach these aims?
a) create a network (watertubes) between thecompanies
b) Eventually install someregeneration unit(cleaning of the water)
It is important to understand that this approach is not limited to
water. It can be applied to vapor, gas, coaling �uids, human
resources...
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 116: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/116.jpg)
Kalundborg (Danemark)
An symbolic example of Industrial eco-park is Kalundborg(Danemark)
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 117: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/117.jpg)
De�nition
What is an � Eco-park � ?
In order to convince companies to participate to the Ecopark, ourmodel should guarantee that:
a) each company will have a lower cost of production in Eco-parkorganization than in stand-alone organization
b) the eco-park organization must generate a lower freshwaterconsumption than with a stand-alone organization
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 118: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/118.jpg)
MOO classical treatment
The Eco-park design was done through Multi-objective Optimizationby the evaluation of Pareto fronts (Gold programming algorithms,scalarization...).
min
Fresh water consumptionIndividual costs of producer 1
...Individual costs of producer n
s.t.
Water balancesTopological constraintsWater quality criteria
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 119: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/119.jpg)
MOO classical treatment
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 120: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/120.jpg)
Alternative approach
The needed change :
. . . to have an independant designer/regulator
. . . to have fair solutions for the companies
Thus we propose to use two di�erent possible models:
Hierarchical optimisation (bi-level optim.)
Nash game concept between the companies
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 121: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/121.jpg)
Single-Leader-Multi-Follower game
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 122: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/122.jpg)
Numerical treatment
This very di�cult problem is treated as follows:
�rst we replace the lower-level (convex) optimization problem bytheir KKT systems; the resulting problem is an MathematicalProgramming with Complementarity Constraints (MPCC);
second the MPCC problem is solved by penalization methods
Numerical results have been obtained with Julia meta-solver coupled
with Gurobi, IPOPT and Baron.
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 123: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/123.jpg)
Single-Leader-Multi-Follower game
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 124: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/124.jpg)
More on Single-Leader-Multi-Followergames
D.A & A. Svensson (J. Optim. Theory Appl. 182 (2019))
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 125: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/125.jpg)
Existence for optimistic SLMF games
Theorem
Assume that F is lower semi-continuous, and for each followeri = 1, ...,M the objective fi is continuous and(x, y−i) 7→ Ci(x, y−i) := {yi | gi(x, y) ≤ 0} is a lower semi-continuousset-valued map which has nonempty compact graph. If the graph ofGNEP is nonempty, then the SLMF game admits an optimisticsolution.
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 126: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/126.jpg)
Example of linear pessimistic SLMF game with nosolutions
Let us consider the SLMFG with two followers
minx∈[0,4]
maxy∈GNEP(x)
−x+ (y1 + y2).
with
miny1 y1
s.t.
y1 ≥ 02y2 − y1 ≤ 2y1 + y2 ≥ x
miny2 y2
s.t.
y2 ≥ 02y1 − y2 ≤ 2y1 + y2 ≥ x
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 127: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/127.jpg)
The solution of the parametric GNEP of the followers is given by
GNEP(x) =
{(0, 0), (2, 2)} if x ≥ 4,{(0, 0)} if x ∈ [0, 4[∅ otherwise.
(5)
Notice that the function
ϕmax(x) := maxy∈GNEP(x)
−x+ (y1 + y2)
=
{0 if x = 4−x if x ∈ [0, 4[
is not lower semi-continuous, so that Weierstrass theorem argument
cannot be applied. And in fact, the value of the problem of the leader
is −4, while there does not exist a point x ∈ [0, 4] with that value.
The pessimistic linear single-leader-two-follower problem has no
optimal solution.
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 128: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/128.jpg)
Going back to applications: IEP
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 129: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/129.jpg)
Another approach: the blind/control input models
In two very recent works we suggested some reformulations of theoptimal design problem:
under some hypothesis (unique process for each company,linearization in the case of regeneration units), we shown thatthe optimal design problem can be reformulated as a classicalMixed Integer Linear Programming problem (MILP);
this problem can be treated with classical tools (CPLEX);
Moreover we inserted a "minimal gain" condition
Costi(xi, xP−i, x
R, E) ≤ αi · STCi, ∀i ∈ IP .
ensuring that each participating company will gain at minimum α%
on its production cost.
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 130: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/130.jpg)
Another appraoch: the blind/control input models
Theorem
For E ∈ E and xR ≥ 0 �xed, the equilibrium set Eq(xR, E) is given by
Eq(xR, E) =
xP
: ∀i ∈ IP ,
zi(x−i) +∑
(k,i)∈Exk,i =
∑(i,j)∈E
xi,j
gi(x−i) ≤ 0zi(x−i) ≥ 0xi∣∣Eci,act
= 0
xi ≥ 0
(6)
Thus, the optimal design problem is equivalent to
minE∈E,x∈R|Emax|
Z(x)
s.t.
x ∈ X,zi(x−i) +
∑(k,i)∈E
xk,i = +∑
(i,j)∈Exi,j , ∀i ∈ I
xi∣∣Eci,act
= 0, ∀i ∈ I
gi(x−i) ≤ 0, ∀i ∈ Izi(x−i) ≥ 0, ∀i ∈ ICosti(xi, x
P−i, x
R, E) ≤ αi · STCi, ∀i ∈ IPx ≥ 0.
(7)
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 131: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/131.jpg)
Some results
11
12 13
36
4 8 15 14 7910521
0
Figure: The con�guration in the case without regeneration units,αi = 0.95 and Coef = 1. Gray nodes are consuming strictly positivefresh water.
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 132: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/132.jpg)
0.75 0.8 0.85 0.9 0.95 1
α
0
1
2
3
4
5
6
7
The n
um
ber
of ente
rprises o
pera
ting s
tand-a
lone
360
370
380
390
400
410
420
430
Glo
bal fr
eshw
ate
r consum
ption
The number of enterprises operating STC
Global Freshwater consumption
Figure: The number of enterprises operating stand-alone and theglobal freshwater consumption with Coef = 1.
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 133: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/133.jpg)
0 2 4 6 8 10 12 14 16 18 20
coef
0
1
2
3
4
5
6
7
Th
e n
um
be
r o
f e
nte
rprise
s o
pe
ratin
g s
tan
d-a
lon
e
365
366
367
368
369
370
371
372
373
374
375
Glo
ba
l fr
esh
wa
ter
co
nsu
mp
tio
n
The number of enterprises operating STC
Global Freshwater consumption
Figure: The number of enterprises operating stand-alone and theglobal freshwater consumption with α = 0.99.
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 134: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/134.jpg)
Some references...
D. Aussel, A. Svensson, Towards Tractable Constraint Quali�cations for Parametric
Optimisation Problems and Applications to Generalised Nash Games, J. Optim.Theory Appl. 182 (2019), 404-416.
D. Aussel, L. Brotcorne, S. Lepaul, L. von Niederhäusern, A Trilevel Model for Best
Response in Energy Demand Side Management, Eur. J. Operations Research 281(2020), 299-315.
D. Aussel, K. Cao Van, D. Salas, Quasi-variational Inequality Problems over Product
sets with Quasimonotone Operators, SIOPT 29 (2019), 1558-1577.
D. Aussel & A. Svensson, Is Pessimistic Bilevel Programming a Special Case of a
Mathematical Program with Complementarity Constraints?, J. Optim. TheoryAppl. 181(2) (2019), 504-520.
D. Aussel & A. Svensson, Some remarks on existence of equilibria, and the validity
of the EPCC reformulation for multi-leader-follower games, J. Nonlinear ConvexAnal. 19 (2018), 1141-1162.
E. Allevi, D. Aussel & R. Riccardi, On a equilibrium problem with complementarity
constraints formulation of pay-as-clear electricity market with demand elasticity,J. Global Optim. 70 (2018), 329-346.
D. Aussel, P. Bendotti and M. Pi²t¥k, Nash Equilibrium in Pay-as-bid Electricity
Market : Part 1 - Existence and Characterisation, Optimization 66:6 (2017),1013-1025.
D. Aussel, P. Bendotti and M. Pi²t¥k, Nash Equilibrium in Pay-as-bid Electricity
Market : Part 2 - Best Response of Producer, Optimization 66:6 (2017), 1027-1053.
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 135: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/135.jpg)
Some references...
M. Ramos, M. Boix, D. Aussel, L. Montastruc, S. Domenech, Water integration in
Eco-Industrial Parks Using a Multi-Leader-Follower Approach, Computers &Chemical Engineering 87 (2016) 190-207.
M. Ramos, M. Boix, D. Aussel, L. Montastruc, S. Domenech, Optimal Design of Water
Exchanges in Eco-Inductrial Parks Through a Game Theory Approach,Computers Aided Chemical Engineering 38 (2016) 1177-1183.
M. Ramos, M. Rocafull, M. Boix, D. Aussel, L. Montastruc & S. Domenech, UtilityNetwork Optimization in Eco-Industrial Parks by a Multi-Leader-Follower game
Methodology, Computers & Chemical Engineering 112 (2018), 132-153.
D. Salas, Cao Van Kien, D. Aussel, L. Montastruc, Optimal design of exchange
networks with blind inputs and its application to Eco-Industrial parks, Computers& Chemical Engineering 143 (2020), 18 pp, published online.
Aussel, K. Cao Van, Control-input approach of of water exchange in Eco-Industrial
Parks, submitted (2020).
Didier AusselBilevel Problems, MPCCs, and Multi-Leader-Follower Games
![Page 136: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/136.jpg)
Quasiconvex optimizationNow the case of GNEP...
MLFG in the setting of quasiconvex optimization
Didier AusselUniv. de Perpignan, France
ALOP autumn school - October 14th, 2020
Coauthors: N. Hadjisavvas (Greece and Saudia), M. Pistek (Czech Republic), Jane Ye (Canada)
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 137: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/137.jpg)
Quasiconvex optimizationNow the case of GNEP...
I - Introduction to quasiconvex optimization
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 138: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/138.jpg)
Quasiconvex optimizationNow the case of GNEP...
Quasiconvexity
A function f : X → IR ∪ {+∞} is said to be quasiconvex on K if,
for all x, y ∈ K and all t ∈ [0, 1],
f (tx + (1− t)y) ≤ max{f (x), f (y)}.
or
for all λ ∈ IR, the sublevel set
Sλ = {x ∈ X : f (x) ≤ λ} is convex.
or
f differentiable
f is quasiconvex ⇐⇒ df is quasimonotone
or
f is quasiconvex ⇐⇒ ∂f is quasimonotone
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 139: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/139.jpg)
Quasiconvex optimizationNow the case of GNEP...
Quasiconvexity
A function f : X → IR ∪ {+∞} is said to be quasiconvex on K if,
for all x, y ∈ K and all t ∈ [0, 1],
f (tx + (1− t)y) ≤ max{f (x), f (y)}.
or
for all λ ∈ IR, the sublevel set
Sλ = {x ∈ X : f (x) ≤ λ} is convex.
or
f differentiable
f is quasiconvex ⇐⇒ df is quasimonotone
or
f is quasiconvex ⇐⇒ ∂f is quasimonotone
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 140: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/140.jpg)
Quasiconvex optimizationNow the case of GNEP...
Quasiconvexity
A function f : X → IR ∪ {+∞} is said to be quasiconvex on K if,
for all x, y ∈ K and all t ∈ [0, 1],
f (tx + (1− t)y) ≤ max{f (x), f (y)}.
or
for all λ ∈ IR, the sublevel set
Sλ = {x ∈ X : f (x) ≤ λ} is convex.
or
f differentiable
f is quasiconvex ⇐⇒ df is quasimonotone
or
f is quasiconvex ⇐⇒ ∂f is quasimonotone
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 141: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/141.jpg)
Quasiconvex optimizationNow the case of GNEP...
Quasiconvexity
A function f : X → IR ∪ {+∞} is said to be quasiconvex on K if,
for all x, y ∈ K and all t ∈ [0, 1],
f (tx + (1− t)y) ≤ max{f (x), f (y)}.
or
for all λ ∈ IR, the sublevel set
Sλ = {x ∈ X : f (x) ≤ λ} is convex.
or
f differentiable
f is quasiconvex ⇐⇒ df is quasimonotone
or
f is quasiconvex ⇐⇒ ∂f is quasimonotone
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 142: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/142.jpg)
Quasiconvex optimizationNow the case of GNEP...
Quasiconvexity
A function f : X → IR ∪ {+∞} is said to be quasiconvex on K if,
for all λ ∈ IR, the sublevel set
Sλ = {x ∈ X : f (x) ≤ λ} is convex.
A function f : X → IR ∪ {+∞} is said to be semistrictly quasiconvexon K if, f is quasiconvex and for any x , y ∈ K ,
f (x) < f (y)⇒ f (z) < f (y), ∀ z ∈ [x , y [.
convex ⇒ semistrictly quasiconvex ⇒ quasiconvex
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 143: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/143.jpg)
Quasiconvex optimizationNow the case of GNEP...
Why not a subdifferential for quasiconvex programming?
No (upper) semicontinuity of ∂f if f is not supposed to beLipschitz
No sufficient optimality condition
x ∈ Sstr (∂f ,C ) =⇒ x ∈ arg minC
fHHH ���
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 144: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/144.jpg)
Quasiconvex optimizationNow the case of GNEP...
Why not a subdifferential for quasiconvex programming?
No (upper) semicontinuity of ∂f if f is not supposed to beLipschitz
No sufficient optimality condition
x ∈ Sstr (∂f ,C ) =⇒ x ∈ arg minC
fHHH ���
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 145: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/145.jpg)
Quasiconvex optimizationNow the case of GNEP...
Why not a subdifferential for quasiconvex programming?
No (upper) semicontinuity of ∂f if f is not supposed to beLipschitz
No sufficient optimality condition
x ∈ Sstr (∂f ,C ) =⇒ x ∈ arg minC
fHHH ���
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 146: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/146.jpg)
Quasiconvex optimizationNow the case of GNEP...
II - Normal approach
a- First definitions
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 147: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/147.jpg)
Quasiconvex optimizationNow the case of GNEP...
A first approach
Sublevel set:
Sλ = {x ∈ X : f (x) ≤ λ}
S>λ = {x ∈ X : f (x) < λ}
Normal operator:
Define Nf (x) : X → 2X∗ by
Nf (x) = N(Sf (x), x)= {x∗ ∈ X ∗ : 〈x∗, y − x〉 ≤ 0, ∀ y ∈ Sf (x)}.
With the corresponding definition for N>f (x)
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 148: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/148.jpg)
Quasiconvex optimizationNow the case of GNEP...
But ...
Nf (x) = N(Sf (x), x) has no upper-semicontinuity properties
N>f (x) = N(S>f (x), x) has no quasimonotonicity properties
ExampleDefine f : R2 → R by
f (a, b) =
{|a| + |b| , if |a| + |b| ≤ 1
1, if |a| + |b| > 1.
Then f is quasiconvex.Consider x = (10, 0), x∗ = (1, 2), y = (0, 10) and y∗ = (2, 1).
We see that x∗ ∈ N<(x) and y∗ ∈ N< (y) (since |a| + |b| < 1 implies (1, 2) · (a − 10, b) ≤ 0 and
(2, 1) · (a, b − 10) ≤ 0) while⟨x∗, y − x
⟩> 0 and
⟨y∗, y − x
⟩< 0. Hence N< is not quasimonotone.
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 149: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/149.jpg)
Quasiconvex optimizationNow the case of GNEP...
But ...
Nf (x) = N(Sf (x), x) has no upper-semicontinuity properties
N>f (x) = N(S>f (x), x) has no quasimonotonicity properties
ExampleDefine f : R2 → R by
f (a, b) =
{|a| + |b| , if |a| + |b| ≤ 1
1, if |a| + |b| > 1.
Then f is quasiconvex.
Consider x = (10, 0), x∗ = (1, 2), y = (0, 10) and y∗ = (2, 1).
We see that x∗ ∈ N<(x) and y∗ ∈ N< (y) (since |a| + |b| < 1 implies (1, 2) · (a − 10, b) ≤ 0 and
(2, 1) · (a, b − 10) ≤ 0) while⟨x∗, y − x
⟩> 0 and
⟨y∗, y − x
⟩< 0. Hence N< is not quasimonotone.
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 150: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/150.jpg)
Quasiconvex optimizationNow the case of GNEP...
But ...
Nf (x) = N(Sf (x), x) has no upper-semicontinuity properties
N>f (x) = N(S>f (x), x) has no quasimonotonicity properties
ExampleDefine f : R2 → R by
f (a, b) =
{|a| + |b| , if |a| + |b| ≤ 1
1, if |a| + |b| > 1.
Then f is quasiconvex.Consider x = (10, 0), x∗ = (1, 2), y = (0, 10) and y∗ = (2, 1).
We see that x∗ ∈ N<(x) and y∗ ∈ N< (y) (since |a| + |b| < 1 implies (1, 2) · (a − 10, b) ≤ 0 and
(2, 1) · (a, b − 10) ≤ 0) while⟨x∗, y − x
⟩> 0 and
⟨y∗, y − x
⟩< 0. Hence N< is not quasimonotone.
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 151: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/151.jpg)
Quasiconvex optimizationNow the case of GNEP...
But ...another example
Nf (x) = N(Sf (x), x) has no upper-semicontinuity properties
N>f (x) = N(S>f (x), x) has no quasimonotonicity properties
Example
Then f is quasiconvex.
We easily see that N(x) is not upper semicontinuous....
These two operators are essentially adapted to the class of semi-strictly
quasiconvex functions. Indeed in this case, for each x ∈ dom f \ arg min f ,
the sets Sf (x) and S<f (x) have the same closure and Nf (x) = N<f (x).
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 152: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/152.jpg)
Quasiconvex optimizationNow the case of GNEP...
But ...another example
Nf (x) = N(Sf (x), x) has no upper-semicontinuity properties
N>f (x) = N(S>f (x), x) has no quasimonotonicity properties
Example
Then f is quasiconvex.
We easily see that N(x) is not upper semicontinuous....
These two operators are essentially adapted to the class of semi-strictly
quasiconvex functions. Indeed in this case, for each x ∈ dom f \ arg min f ,
the sets Sf (x) and S<f (x) have the same closure and Nf (x) = N<f (x).
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 153: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/153.jpg)
Quasiconvex optimizationNow the case of GNEP...
But ...another example
Nf (x) = N(Sf (x), x) has no upper-semicontinuity properties
N>f (x) = N(S>f (x), x) has no quasimonotonicity properties
Example
Then f is quasiconvex.
We easily see that N(x) is not upper semicontinuous....
These two operators are essentially adapted to the class of semi-strictly
quasiconvex functions. Indeed in this case, for each x ∈ dom f \ arg min f ,
the sets Sf (x) and S<f (x) have the same closure and Nf (x) = N<f (x).Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 154: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/154.jpg)
Quasiconvex optimizationNow the case of GNEP...
II - Normal approach
b- Adjusted sublevel setsand
normal operator
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 155: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/155.jpg)
Quasiconvex optimizationNow the case of GNEP...
Definition
Adjusted sublevel set
For any x ∈ dom f , we define
Saf (x) = Sf (x) ∩ B(S<f (x), ρx)
where ρx = dist(x ,S<f (x)), if S<f (x) 6= ∅
and Saf (x) = Sf (x) if S<f (x) = ∅.
Saf (x) coincides with Sf (x) if cl(S>f (x)) = Sf (x)
e.g. f is semistrictly quasiconvex
Proposition
Let f : X → IR ∪ {+∞} be any function, with domain dom f . Then
f is quasiconvex ⇐⇒ Saf (x) is convex ,∀ x ∈ dom f .
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 156: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/156.jpg)
Quasiconvex optimizationNow the case of GNEP...
Definition
Adjusted sublevel set
For any x ∈ dom f , we define
Saf (x) = Sf (x) ∩ B(S<f (x), ρx)
where ρx = dist(x ,S<f (x)), if S<f (x) 6= ∅
and Saf (x) = Sf (x) if S<f (x) = ∅.
Saf (x) coincides with Sf (x) if cl(S>f (x)) = Sf (x)
e.g. f is semistrictly quasiconvex
Proposition
Let f : X → IR ∪ {+∞} be any function, with domain dom f . Then
f is quasiconvex ⇐⇒ Saf (x) is convex ,∀ x ∈ dom f .
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 157: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/157.jpg)
Quasiconvex optimizationNow the case of GNEP...
Definition
Adjusted sublevel set
For any x ∈ dom f , we define
Saf (x) = Sf (x) ∩ B(S<f (x), ρx)
where ρx = dist(x ,S<f (x)), if S<f (x) 6= ∅
and Saf (x) = Sf (x) if S<f (x) = ∅.
Saf (x) coincides with Sf (x) if cl(S>f (x)) = Sf (x)
e.g. f is semistrictly quasiconvex
Proposition
Let f : X → IR ∪ {+∞} be any function, with domain dom f . Then
f is quasiconvex ⇐⇒ Saf (x) is convex ,∀ x ∈ dom f .
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 158: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/158.jpg)
Quasiconvex optimizationNow the case of GNEP...
Definition
Adjusted sublevel set
For any x ∈ dom f , we define
Saf (x) = Sf (x) ∩ B(S<f (x), ρx)
where ρx = dist(x ,S<f (x)), if S<f (x) 6= ∅
and Saf (x) = Sf (x) if S<f (x) = ∅.
Saf (x) coincides with Sf (x) if cl(S>f (x)) = Sf (x)
e.g. f is semistrictly quasiconvex
Proposition
Let f : X → IR ∪ {+∞} be any function, with domain dom f . Then
f is quasiconvex ⇐⇒ Saf (x) is convex ,∀ x ∈ dom f .
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 159: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/159.jpg)
Quasiconvex optimizationNow the case of GNEP...
Adjusted normal operator
Adjusted sublevel set:
For any x ∈ dom f , we define
Saf (x) = Sf (x) ∩ B(S<f (x), ρx)
where ρx = dist(x ,S<f (x)), if S<f (x) 6= ∅.
Ajusted normal operator:
Naf (x) = {x∗ ∈ X ∗ : 〈x∗, y − x〉 ≤ 0, ∀ y ∈ Sa
f (x)}
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 160: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/160.jpg)
Quasiconvex optimizationNow the case of GNEP...
Example
x
��
��
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 161: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/161.jpg)
Quasiconvex optimizationNow the case of GNEP...
Example
x
��
��
B(S<f (x), ρx)
Saf (x) = Sf (x) ∩ B(S<
f (x), ρx)
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 162: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/162.jpg)
Quasiconvex optimizationNow the case of GNEP...
Example
Saf (x) = Sf (x) ∩ B(S<
f (x), ρx)
Naf (x) = {x∗ ∈ X ∗ : 〈x∗, y − x〉 ≤ 0, ∀ y ∈ Sa
f (x)}
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 163: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/163.jpg)
Quasiconvex optimizationNow the case of GNEP...
An exercice.........
Let us draw the normal operator value Naf (x , y) at the points
(x , y) = (0.5, 0.5), (x , y) = (0, 1), (x , y) = (1, 0), (x , y) = (1, 2),(x , y) = (1.5, 0) and (x , y) = (0.5, 2).
Operator Naf provide information at any point!!!
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 164: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/164.jpg)
Quasiconvex optimizationNow the case of GNEP...
An exercice.........
Let us draw the normal operator value Naf (x , y) at the points
(x , y) = (0.5, 0.5), (x , y) = (0, 1), (x , y) = (1, 0), (x , y) = (1, 2),(x , y) = (1.5, 0) and (x , y) = (0.5, 2).
Operator Naf provide information at any point!!!
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 165: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/165.jpg)
Quasiconvex optimizationNow the case of GNEP...
Basic properties of Naf
Nonemptyness:
Proposition
Let f : X → IR ∪ {+∞} be lsc. Assume that rad. continuous on dom for dom f is convex and intSλ 6= ∅, ∀λ > infX f . Then
f is quasiconvex ⇔ Naf (x) \ {0} 6= ∅, ∀ x ∈ dom f \ arg min f .
Quasimonotonicity:
The normal operator Naf is always quasimonotone
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 166: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/166.jpg)
Quasiconvex optimizationNow the case of GNEP...
Upper sign-continuity
• T : X → 2X∗ is said to be upper sign-continuous on K iff for anyx , y ∈ K , one have :
∀ t ∈ ]0, 1[, infx∗∈T (xt)
〈x∗, y − x〉 ≥ 0
=⇒ supx∗∈T (x)
〈x∗, y − x〉 ≥ 0
where xt = (1− t)x + ty .
upper semi-continuous⇓
upper hemicontinuous⇓
upper sign-continuous
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 167: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/167.jpg)
Quasiconvex optimizationNow the case of GNEP...
locally upper sign continuity
Definition
Let T : K → 2X∗ be a set-valued map.
T is called locally upper sign-continuous on K if, for any x ∈ K thereexist a neigh. Vx of x and a upper sign-continuous set-valued mapΦx(·) : Vx → 2X∗ with nonempty convex w∗-compact values such thatΦx(y) ⊆ T (y) \ {0}, ∀ y ∈ Vx
Continuity of normal operator
Proposition
Let f be lsc quasiconvex function such that int(Sλ) 6= ∅, ∀λ > inf f .
Then Naf is locally upper sign-continuous on dom f \ arg min f .
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 168: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/168.jpg)
Quasiconvex optimizationNow the case of GNEP...
locally upper sign continuity
Definition
Let T : K → 2X∗ be a set-valued map.
T is called locally upper sign-continuous on K if, for any x ∈ K thereexist a neigh. Vx of x and a upper sign-continuous set-valued mapΦx(·) : Vx → 2X∗ with nonempty convex w∗-compact values such thatΦx(y) ⊆ T (y) \ {0}, ∀ y ∈ Vx
Continuity of normal operator
Proposition
Let f be lsc quasiconvex function such that int(Sλ) 6= ∅, ∀λ > inf f .
Then Naf is locally upper sign-continuous on dom f \ arg min f .
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 169: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/169.jpg)
Quasiconvex optimizationNow the case of GNEP...
III
Quasiconvex programming
a- Optimality conditions
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 170: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/170.jpg)
Quasiconvex optimizationNow the case of GNEP...
Quasiconvex programming
Let f : X → IR ∪ {+∞} and K ⊆ dom f be a convex subset.
(P) find x ∈ K : f (x) = infx∈K
f (x)
Perfect case: f convex
f : X → IR ∪ {+∞} a proper convex function
K a nonempty convex subset of X , x ∈ K + C.Q.
Thenf (x) = inf
x∈Kf (x) ⇐⇒ x ∈ Sstr (∂f ,K )
What about f quasiconvex case?
x ∈ Sstr (∂f (x),K ) =⇒ x ∈ arg minK
fHHH ���
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 171: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/171.jpg)
Quasiconvex optimizationNow the case of GNEP...
Quasiconvex programming
Let f : X → IR ∪ {+∞} and K ⊆ dom f be a convex subset.
(P) find x ∈ K : f (x) = infx∈K
f (x)
Perfect case: f convex
f : X → IR ∪ {+∞} a proper convex function
K a nonempty convex subset of X , x ∈ K + C.Q.
Thenf (x) = inf
x∈Kf (x) ⇐⇒ x ∈ Sstr (∂f ,K )
What about f quasiconvex case?
x ∈ Sstr (∂f (x),K ) =⇒ x ∈ arg minK
fHHH ���
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 172: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/172.jpg)
Quasiconvex optimizationNow the case of GNEP...
Quasiconvex programming
Let f : X → IR ∪ {+∞} and K ⊆ dom f be a convex subset.
(P) find x ∈ K : f (x) = infx∈K
f (x)
Perfect case: f convex
f : X → IR ∪ {+∞} a proper convex function
K a nonempty convex subset of X , x ∈ K + C.Q.
Thenf (x) = inf
x∈Kf (x) ⇐⇒ x ∈ Sstr (∂f ,K )
What about f quasiconvex case?
x ∈ Sstr (∂f (x),K ) =⇒ x ∈ arg minK
fHHH ���
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 173: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/173.jpg)
Quasiconvex optimizationNow the case of GNEP...
Sufficient optimality condition
Theorem
f : X → IR ∪ {+∞} quasiconvex, radially cont. on dom f
C ⊆ X such that conv(C ) ⊂ dom f .Suppose that C ⊂ int(dom f ) or AffC = X .
Then x ∈ S(Naf \ {0},C ) =⇒ ∀ x ∈ C , f (x) ≤ f (x).
where x ∈ S(Naf \ {0},K) means that there exists x∗ ∈ Na
f (x) \ {0} such that
〈x∗, c − x〉 ≥ 0, ∀ c ∈ C .
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 174: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/174.jpg)
Quasiconvex optimizationNow the case of GNEP...
Necessary and Sufficient conditions
Proposition
Let C be a closed convex subset of X , x ∈ C and f : X → IR becontinuous semistrictly quasiconvex such that int(Sa
f (x)) 6= ∅ andf (x) > infX f .Then the following assertions are equivalent:
i) f (x) = minC f
ii) x ∈ Sstr (Naf \ {0},C )
iii) 0 ∈ Naf (x) \ {0}+ NK (C , x).
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 175: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/175.jpg)
Quasiconvex optimizationNow the case of GNEP...
GNEP reformulation in quasiconvex case
To simplify the notations, we will denote, for any i and any x ∈ Rn, by Si (x)and Ai (x
−i ) the subsets of Rni
Si (x) = Saθi (·,x−i )(x
i ) and Ai (x−i ) = arg min
Rniθi (·, x−i ).
In order to construct the variational inequality problem we define the followingset-valued map Na
θ : Rn → 2Rn
which is described,
for any x = (x1, . . . , xp) ∈ Rn1 × . . .× Rnp , by
Naθ(x) = F1(x)× . . .× Fp(x),
where Fi (x) =
{B i (0, 1) if x i ∈ Ai (x
−i )
co(Naθi
(x i ) ∩ Si (0, 1)) otherwise
The set-valued map Naθ has nonempty convex compact values.
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 176: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/176.jpg)
Quasiconvex optimizationNow the case of GNEP...
Sufficient condition
In the following we assume that X is a given nonempty subset X of Rn , such that for any i , the set Xi (x−i ) is
given as
Xi (x−i ) = {x i ∈ Rni : (x i , x−i ) ∈ X}.
Theorem
Let us assume that, for any i , the function θi is continuous andquasiconvex with respect to the i-th variable. Then every solution ofS(Na
θ ,X ) is a solution of the GNEP.
Note that the link between GNEP and variational inequality is valid even
if the constraint set X is neither convex nor compact.
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 177: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/177.jpg)
Quasiconvex optimizationNow the case of GNEP...
Sufficient condition
In the following we assume that X is a given nonempty subset X of Rn , such that for any i , the set Xi (x−i ) is
given as
Xi (x−i ) = {x i ∈ Rni : (x i , x−i ) ∈ X}.
Theorem
Let us assume that, for any i , the function θi is continuous andquasiconvex with respect to the i-th variable. Then every solution ofS(Na
θ ,X ) is a solution of the GNEP.
Note that the link between GNEP and variational inequality is valid even
if the constraint set X is neither convex nor compact.
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 178: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/178.jpg)
Quasiconvex optimizationNow the case of GNEP...
Lemma
Let i ∈ {1, . . . , p}. If the function θi is continuous quasiconvex withrespect to the i-th variable, then,
0 ∈ Fi (x) ⇐⇒ x i ∈ Ai (x−i ).
Proof. It is sufficient to consider the case of a point x such that x i 6∈ Ai (x−i ). Since θi (·, x−i ) is continuous at
x i , the interior of Si (x) is nonempty. Let us denote by Ki the convex cone
Ki = Naθi
(x i ) = (Si (x)− x i )◦.
By quasiconvexity of θi , Ki is not reduced to {0}. Let us first observe that, since Si (x) has a nonempty interior,Ki is a pointed cone, that is Ki ∩ (−Ki ) = {0}.Now let us suppose that 0 ∈ Fi (x). By Caratheodory theorem, there exist vectors vi ∈ [Ki ∩ Si (0, 1)],i = 1, . . . , n + 1 and scalars λi ≥ 0, i = 1, . . . , n + 1 with
n+1∑i=1
λi = 1 and 0 =n+1∑i=1
λi vi .
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 179: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/179.jpg)
Quasiconvex optimizationNow the case of GNEP...
Since there exists at least one r ∈ {1, . . . , n + 1} such that λr > 0 we have
vr = −n+1∑
i=1,i 6=r
λi
λrvi
which clearly shows that vr is an element of the convex cone −Ki . But vr ∈ Si (0, 1) and thus vr 6= 0. Thiscontradicts the fact that Ki is pointed and the proof is complete.
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 180: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/180.jpg)
Quasiconvex optimizationNow the case of GNEP...
Proof of necessary condition
Proof. Let us consider x to be a solution of S(Naθ, X ). There exists v ∈ Na
θ(x) such that
〈v, y − x〉 ≥ 0, ∀ y ∈ X . (∗)
Let i ∈ {1, . . . , p}.
If x i ∈ Ai (x−i ) then obviously x i ∈ Soli (x
−i ).
Otherwise v i ∈ Fi (x) = co(Naθi
(x i ) ∩ Si (0, 1)). Thus, according to Lemma 2, there exist λ > 0 and
ui ∈ Naθi
(x i ) \ {0} satisfying v i = λui .
Now for any x i ∈ Xi (x−i ), consider y =
(x1, . . . , x i−1, x i , x i+1, . . . , xp
).
From (∗) one immediately obtains that 〈ui , x i − x i 〉 ≥ 0. Since x i is an arbitrary element of Xi (x−i ), we have
that x i is a solution of S(Naθi\ {0}, Xi (x
−i )) and therefore, according to Prop. 4,
x i ∈ Soli (x−i )
Since i was arbitrarily chosen we conclude that x solves the GNEP.
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization
![Page 181: Bilevel Problems, MPCCs, and Multi-Leader-Follower Games · 2020. 10. 15. · Didier Aussel Lab. Promes UPR CNRS 8521, University of Perpignan, ranceF ALOP autumn school - October](https://reader035.vdocuments.us/reader035/viewer/2022071608/61464a2c8f9ff81254202b89/html5/thumbnails/181.jpg)
Quasiconvex optimizationNow the case of GNEP...
Necessary and sufficient condition
Theorem
Let us suppose that, for any i , the loss function θi is continuous andsemistrictly quasiconvex with respect to the i-th variable. Further assumethat the set X is a nonempty convex subset of RN . Then
any solution of the variational inequality S(Naθ ,X ) is a solution
of the GNEP
and
any solution of the GNEP is a solution of the quasi-variationalinequality QVI (Na
θ ,X )
where X stands for the set-valued map defined on R2 by
X (x) =
p∏i=1
Xi (x−i )
.
D.A. & J. Dutta, Oper. Res. Letters, 2008.
Didier Aussel Univ. de Perpignan, France MLFG in the setting of quasiconvex optimization