Séminaire OrléansA non-parametric sampling approach for nonlinear stochastic programming
Thursday 24 May 2007 14:00 - Orléans - Salle S104 (Bât Sciences)
A major difficulty in stochastic programming is the choice of probability distributions present in the objective functions. This is especially true for parameters estimation problems, as maximum likelihood or least-squares. We propose here to focus on the cumulating functions: sampling over [0,1], we estimate inverse cumulating functions by means of B-cubic splines. The resulting optimization problem is a nonlinear program subject to additional monotonicity constraints. This problem is solved with an adapted constrained trust-region procedure, using projections to keep the iterates feasible.