Groupe d’études et de recherche en analyse des décisions

Some recent work in derivative free optimization

Andrew R. Conn IBM TJ Watson Research Center, États-Unis

Derivative free methods are amongst the most commonly used optimization techniques in practise but it is absolutely essential that they balance the geomtry of sample points with efficiency. I will present a general framework for a trust region derivative free algorithm that, for the models, assumes only that they can be made to satisfy Taylor-like error bounds. The results have broader implications (for example to, second-order trust-region methods and the connection between geometry and a (scaled) basis matrix condition number) and are of both theoretical and practical interest.

Joint work with Katya Scheinberg and Luis Vicente