28.05.2013 Views

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 ...

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

tel-00703797, version 2 - 7 Jun 2012<br />

In the jointly convex case, we get<br />

3.2. Methods to solve <strong>non</strong>linear equations<br />

ˆV (x) = 0 and x ∈ X(x), (3.11)<br />

where the set X(x) = {y ∈ R n , g(yi, x−i) ≤ 0}. Still the computation of ˆ V is a complex optimization<br />

over a constrained set X(x). As in the previous subsection, the class of GNE called<br />

variational equilibrium can be characterized by the NI formulation. We have the folllowing<br />

theorem.<br />

Theorem. Assuming θi and g are C 1 functions and g is convex and θi player-convex. x ⋆ is<br />

a variational equilibrium if and only if x ⋆ ∈ X and V (x ⋆ ) = 0 with V <strong>de</strong>fined as<br />

V (x) = sup ψ(x, y).<br />

y∈X<br />

In the rest of the paper, we do not study all algorithms but rather focus on the most<br />

promising ones. We restrict our attention to general GNEPs and algorithms to solve the KKT<br />

system presented in Subsection 3.1.1. So, we do not study jointly convex GNEPs for which<br />

special methods have been proposed in the literature. These two situations differs wi<strong>de</strong>ly,<br />

since in the general GNEP, we have to solve a <strong>non</strong>linear equation, while for the jointly convex<br />

case, we solve a fixed point equation or a minimization problem.<br />

3.2 Methods to solve <strong>non</strong>linear equations<br />

As introduced in many optimization books, see, e.g., Dennis and Schnabel (1996); Nocedal<br />

and Wright (2006); Bonnans et al. (2006), an optimization method to solve a <strong>non</strong>linear<br />

equation or more generally to find the minimum of a function is ma<strong>de</strong> of two components:<br />

a local method and a globalization scheme. Assuming the initial point is not “far” from the<br />

root or the optimal point, local methods use a local approximation of the function, generally<br />

linear or quadratic approximation based on the Taylor expansion, that is easier to solve. The<br />

globalization studies adjustments to be carried out, so that the iterate sequence still converges<br />

when algorithms are badly initialized.<br />

To emphasize the prominent role of the globalization, we first look at a simple example of<br />

a <strong>non</strong>linear equation. Let F : R2 ↦→ R2 be <strong>de</strong>fined as<br />

<br />

x2 F (x) = 1 + x2 2<br />

− 2<br />

e x1−1 + x 3 2 − 2<br />

This function only has two roots x ⋆ = (1, 1) and ¯x = (−0.7137474, 1.2208868). We notice that<br />

the second component of F explo<strong><strong>de</strong>s</strong> as x1 tends to infinity.<br />

On Figure 3.2, we plot the contour level of the norm ||F (x)||2, as well as two iterate<br />

sequences (xn), (yn) (see numbers 0, 1, 2,. . . ) starting from the point (x0, y0) = (−1, −3/2).<br />

The first sequence (xn) corresponds to a “pure” Newton method, which we will present after,<br />

whereas the second sequence (yn) combine the Newton method with a line search (LS). We<br />

can observe the sequence (yn) converges less abruptly to the solution x ⋆ than the sequence<br />

(xn).<br />

On Figure 3.3, we plot the contour level of the norm ||F (x)||2 with two iterate sequences<br />

(xn), (yn), for pure and line-search Newton, respectively. But this time, sequences are initiated<br />

at (x0, y0) = (2, 1/2). Despite being close the solution ¯x, the pure sequence (xn) wan<strong>de</strong>rs in<br />

<br />

.<br />

143

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!