Etude des marchés d'assurance non-vie à l'aide d'équilibres de ...
Etude des marchés d'assurance non-vie à l'aide d'équilibres de ...
Etude des marchés d'assurance non-vie à l'aide d'équilibres de ...
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
tel-00703797, version 2 - 7 Jun 2012<br />
Chapitre 3. Calcul d’équilibre <strong>de</strong> Nash généralisé<br />
where u = (x, λ, w) ∈ Rn × Rm ⋆+ × Rm ⋆+ and ζ > m. This reformulation of the potential<br />
function emphasizes the three components u = (x, λ, w). For the line-search, the gradient ∇p<br />
is given by<br />
⎛<br />
⎞<br />
∇p(x, λ, w) =<br />
⎜<br />
⎝<br />
2ζ<br />
||x|| 2 2 +||λ||2 2 +||w||2 2<br />
2ζ<br />
||x|| 2 2 +||λ||2 2 +||w||2 2<br />
2ζ<br />
||x|| 2 2 +||λ||2 2 +||w||2 2<br />
x<br />
λ − λ−1 w − w−1 ⎟<br />
⎠ ,<br />
where λ and w have positive components and terms λ −1 and w −1 correspond to the componentwise<br />
inverse vector. Compared to the semismooth reformulation, the root function H is now<br />
C1 . The Jacobian is given by<br />
⎛<br />
Jac xD(x, λ)<br />
JacH(x, λ, w) = ⎝<br />
diag ∇xigi (x) <br />
Jac xg(x)<br />
<br />
i<br />
0<br />
0<br />
I<br />
⎞<br />
⎠ .<br />
0 diag[w] diag[λ]<br />
As reported in Dreves et al. (2011), the computation of the direction dk = (dx,k, dλ,k, dw,k) in<br />
Equation (3.30) can be simplified due to the special structure of the above Jacobian matrix.<br />
The system reduces to a linear system of n equations to find dx,k and the 2m components<br />
dλ,k, dw,k are simple linear algebra. Using the classic chain rule, the gradient of the merit<br />
function is given by<br />
∇ψ(x, λ, w) = JacH(x, λ, w) T ∇p(H(x, λ, w)).<br />
Again the computation of this gradient can be simplified due to the sparse structure of JacH.<br />
Theorem 4.3 of Dreves et al. (2011) is the direct application of the previous theorem in the<br />
GNEP context. We do not restate here their theorem, but present their <strong>non</strong>singularity result<br />
given in Theorem 4.6. The Jacobian matrix is <strong>non</strong>singular, if the matrix Jac xD(x, λ) is<br />
<strong>non</strong>singular and<br />
M = Jac xg(x)Jac xD(x, λ) −1 diag ∇xigi (x) <br />
(3.31)<br />
is a P0-matrix. This is exactly Equation (3.28) given in the semismooth setting.<br />
3.3 Numerical results<br />
In this section, we perform a numerical illustration to compare the different methods<br />
presented in this paper. The implementation is done in the R statistical software and the<br />
package GNE, freely available on internet.<br />
Our test problem is a simple two-player polynomial-objective game for which there are<br />
four generalized Nash equilibria. The objective functions (to be minimized) are given by<br />
θ1(x) = (x1 − 2) 2 (x2 − 4) 4 and θ2(x) = (x2 − 3) 2 (x1) 4 ,<br />
for x ∈ R 2 , while the constraint functions are<br />
g1(x) = x1 + x2 − 1 ≤ 0 and g2(x) = 2x1 + x2 − 2 ≤ 0.<br />
Objective functions are player strictly convave. This problem is simple but not simplistic,<br />
since second-or<strong>de</strong>r partial <strong>de</strong>rivatives of objective functions are not constant, as for other<br />
162<br />
i