Retour

Numerical optimization

ECTS : 4

Volume horaire : 48

Description du contenu de l'enseignement :

Numerical Optimisation

1. Introduction : a review of basic concepts in optimisation

(a) Optimality conditions, algorithms, convergence rates.

2. First part : Unconstrained optimisation-deterministic methods

(a) A crash course on gradient descent for smooth functions.

(b) The link with gradient flows.

(c) The case of non-convex functions.

(d) Acceleration of gradient descents.

(e) Newton and quasi-Newton methods.

(f) Complement : Back-propagation and machine learning.

3. Second part : Constrained optimisation-deterministic methods

(a) Penalisation method.

(b) The projected gradient method.

(c) Lagrange multipliers and duality-the interior point method.

4. Third part : Unconstrained optimisation-an introduction to stochastic methods

(a) Basic concepts in stochastic gradient descent. Convergence of the algorithm.

(b) Acceleration of stochastic gradient descent.

(c) (Mini)Batches.

Compétence à acquérir :

Mastering traditional techniques in numerical optimisation.

Université Paris Dauphine - PSL - Place du Maréchal de Lattre de Tassigny - 75775 PARIS Cedex 16 - 21/11/2024