Surrogate Assisted Optimization Methods: Recent Developments and Challenges
published: July 20, 2009, recorded: July 2009, views: 624
Report a problem or upload filesIf you have found a problem with this lecture or would like to send us extra material, articles, exercises, etc., please use our ticket system to describe your request and upload the data.
Enter your e-mail into the 'Cc' field, and we will keep you updated with your request's status.
There is an ever increasing trend in the use of more and more accurate and often more computationally expensive analyses tools in early stages of design. Optimization using such computationally expensive analyses tools demand the use of surrogate assisted methods, where a surrogate or an approximation is used in lieu of the expensive analysis to contain the computational time within affordable limits. The performance of such methods is known to be largely dependent on the choice of the underlying optimization algorithm, the surrogate model, the training and surrogate model management schemes. This presentation will introduce a surrogate assisted optimization framework developed by the MDO Group at UNSW@ADFA which alleviates some of the common problems associated with the current approaches. In the proposed approach, surrogates of multiple types (MLP, RBF, Kriging and RSM) coexist within the optimization framework at all times and the surrogate with the least prediction error (based on neighborhood RMSE) is used to approximate the objective and the constraint functions individually. An external archive of all solutions evaluated via actual analysis is maintained to train the surrogates, while a surrogate validity check is performed prior to its use to avoid misguiding the search (in the event of poor approximation or attempts to approximate in unexplored regions). The underlying optimization algorithm is a population based, elitist evolutionary algorithm which explicitly maintains marginally infeasible solutions for a faster rate of convergence. Apart from standard recombination schemes used in any evolutionary algorithm, a memetic recombination operator is embedded to further improve the rate of convergence. A number of examples will be presented to illustrate the performance of the proposed schemes. Finally, the presentation will list areas of further development and present some preliminary results of constrained many objective optimization and spatial approximation schemes that are currently being developed by the group.
Link this pageWould you like to put a link to this lecture on your homepage?
Go ahead! Copy the HTML snippet !