| 1. |
- Holmberg, Joakim, et al.
(författare)
-
VerOpt - MATLAB Driven Versatile Optimization
- 2001
-
Ingår i: Program and proceedings for the Nordic MATLAB Conference Oslo, October 17-18, 2001. - Trondheim : Comsol. - 8299595509 ; , s. 207-212
-
Konferensbidrag (övrigt vetenskapligt/konstnärligt)abstract
- VerOpt, a MATLAB driven versatile optimization environment, enables the choice of a suitable optimization routine, parallelization over TCP/IP and the use of external solvers. VerOpt is the result of working towards the creation of a versatile yet effective environment for applied optimization studies. This paper presents the concepts behind VerOpt, including how and why we use parallelization, and the lessons learnt when using external solvers. The paper also gives a comparison of implemented optimization routines when applied to test problems. Currently, links to three external solvers are implemented. Two of them come from the commercial software market for engineering solutions: ANSYS (version 5.6 University High), a general purpose FE-code and C-MOLD (version 2000.7.1), a code for injection molding. The third solver is from the academic world, AnyBody, a code for biomechanical studies. The implemented optimization routines referred to are Method of Moving Asymptotes (MMA), Simulated Annealing (SA) and a genetic algorithm (GA). The MMA is a gradient-based algorithm whereas the other two can be classified as stochastic. The results of the comparison of the implemented optimization routines, in which �fmincon� from the MATLAB Optimization Toolbox is also used, show that MMA is generally the fastest routine, but does not always find the best solution. However, in test cases when parallelization is used the comparison is not ideal, since the parallelization procedures for the algorithms are not equivalent. When optimization routines are based on numerically computed gradients, such as MMA, they are embarrassingly parallel. This is because the gradients are independent of each other, which makes it possible to compute them simultaneously, but on different processors. For a stochastic routine such as SA a different approach is needed. In our case we have used a simple form of domain decomposition. An interesting result is that, in the test case involving ANSYS, we found that using ANSYS alone, as solver as well as optimizer, did not give such a good solution as using VerOpt. A clear future development is to add a greater number of different types of optimization routines. A possible future development is to transform VerOpt into something that is more akin to a regular style MATLAB Toolbox. Irrespective of this development, VerOpt will be a significant aid for education as well as research in applied optimization. It will also serve the authors as the environment for further research in the fields of injection molding and biomechanics.
|
|
