A derivative-free algorithm for sparse unconstrained optimization problems

We consider the problem of minimizing a function whose derivatives are not available. This paper first presents an algorithm for solving problems of this class using interpolation polynomials and trust-region techniques. We then show how both the data structure and the procedure allowing to build the interpolating polynomials may be adapted in a suitable way to consider problems for which the Hessian matrix is known to be sparse with a general sparsity pattern. The favourable behaviour of the resulting algorithm is confirmed with numerical experiments illustrating the advantages of the method in terms of storage, speed and function evaluations, the latter criterion being particularly important in the framework of derivative-free optimization.