Loading...
Please wait, while we are loading the content...
Similar Documents
Accelerating Aerodynamic Shape Design Using Metamodel-assisted Particle Swarm Optimization
| Content Provider | Semantic Scholar |
|---|---|
| Copyright Year | 2007 |
| Abstract | Particle Swarm Optimization (PSO) [1] has emerged as a very robust tool fo r solving practical optimization problems. Compared to evolutionary algorithms, PSO has smaller number o f parameters and usually converges faster. However it still requires a large number of fu nction evaluations in order to reliably locate an optimum solution. With the increasing use of high fidelity analysis tool (like CFD) in optimization, the cost of one function evaluation can be rather high. In the c as of evolutionary algorithms this problem has been solved to some extent by the use of metamodels (als o called surrogate models). A metamodel (literally, a model of a model) provides a cheaper alterna tiv to the expensive function. The most commonly used type of metamodels are function approximation s like response surfaces, neural networks, radial basis functions (RBF) and kriging. D ue to their ability to model highly non-linear multi-dimensional functions using sparse datasets, RBF and krig ing have emerged as the preferred tool for metamodel-assisted optimization. The high dimensionality and complex function landscapes encountered in pra ctical design problems precludes the use of global metamodels. Global approximations are also cos tly to construct; this is particularly so due to the need to iteratively update the model as new function value s re obtained during the optimization process. Local metamodels on the other hand can be construc ted easily and cheaply, and can better model the complicated functions encountered in engineering d sign. Giannakoglou et al. [2] have proposed a two-level evaluation strategy, called Inexact P re-Evaluation (IPE), to reduce the computational time related to GAs. It relies on the observation that numerous co st function evaluations are useless, since numerous individuals do not survive to the selection o perat r. Hence, it is not necessary to determine their fitness accurately. The strategy proposed by Gia nnakoglou consists in using local metamodels to pre-evaluate the fitness of the individuals in the population. The only a small portion of the population which corresponds to the most promising individuals are elected (pre-screening) and accurately evaluated using the original and expensive model. Inspired by the success of GAs combined with metamodels and IPE, we study th e application of a similar strategy to particle swarm optimization. Like genetic algorithms, PSO is also a r ank-based algorithm; the actual magnitude of cost function of each particle is not importan t but only their relative 0 500 |
| File Format | PDF HTM / HTML |
| Alternate Webpage(s) | http://math.tifrbng.res.in/~praveen/doc/pso-eccomas.pdf |
| Language | English |
| Access Restriction | Open |
| Content Type | Text |
| Resource Type | Article |
