Search

It would be far too expensive to determine a search direction using the pure form of Newton's method, by computing the Hessian matrix of second partial derivatives of the problem functions. (This would roughly square the number of spreadsheet recalculations required to solve the problem.) Instead, an direction is chosen through an estimation method. The default choice Newton uses a quasi-Newton (or BFGS) method, which maintains an approximation to the Hessian matrix; this requires more storage (an amount proportional to the square of the number of currently binding constraints) but performs very well in practice. The alternative choice Conjugate uses a conjugate gradient method, which does not require storage for the Hessian matrix and still performs well in most cases. The choice you make here is not crucial, since the GRG solver is capable of switching automatically between the quasi-Newton and conjugate gradient methods depending on the available storage.