Trust-region based algorithms for singular nonlinear optimization with applications to discrete choice models