A Rigorous framework for optimisation using surrogate objectives
John DennisRice University
Thursday, May 25th, 3.00 p.m. CERFACS Conference Room
Abstract :
Some important problems in computational engineering involve the design or control of products or processes governed by simulations that may cost upwards of a week of high performance workstation time for one run. Even then, it is common for a significant percentage of such expensive runs to produce no value. We will give example problems from design, both for maintainability and for performance as well as an example from manufacturing process control.
Standard practice in engineering is to attack such problems by a one-shot approach that involves fitting interpolatory models to data obtained by running the expensive simulations at carefully chosen points in decision space. This surface is used as an inexpensive surrogate for the true objective function in the optimization process. Sometimes simulations that incorporate less than full physics in the PDE's are used as surrogates, as are neural nets.
The standard one-shot approach is completely ad hoc, mistrusted by engineers, and it often fails to improve the nominal design. We will outline a new approach developed with industrial collaboration. This approach is supported by a rigorous convergence theory, and it works great in the examples so far.
This work is joint with colleagues Charles Audet and Doug Moore at Rice.
Cerfacs' Conferences 1999-2001 Home Page
