RBDO Using PCE Metamodel: Optimal Design Changes with Different Optimization Bounds

UQ with UQLab Community Q&A and How To
, ,
1

Hi everyone,

I appreciate the UQLab software and the literature resources available. Thank you very much.

I have a question regarding the optimal design point in RBDO using PCE metamodels.

First condition: My code runs with an optimization bound of [0.010; 0.020], and after the analysis, it finds an optimal design of 1.2741e-02.

Second condition: When the same code runs with an optimization bound of [0.010; 0.013], the program converges to an optimal design of 1.2555e-02.

My question is: Is the first condition finding the true optimal design? Shouldn’t both conditions converge to the same optimal design point?

The code I am using is as follows:

  • RBDO code: Z_B_uq_Rotor_Laval_Surrogate_PCE_RBDO

Z_B_uq_Rotor_Laval_Surrogate_PCE_RBDO.m (6.1 KB)

  • Codes to run the limit state function:
    Archive.zip (16.1 KB)

Thank you very much for all your support.

2

Hi @Barbara_Oliveira_Men

It is quite possible that they will not converge to the exact same design point. Some reasons for this can be

  • The CCMAES algorithm has a random component, so different initializations may produce different results, even if you specify the random seed.
  • The algorithm may stop too early because it has too soft stopping criteria.
  • The function you are optimizing has a complex topology with possibly many local minima. The solver may end up in different minima.

Best regards
Styfen

3

Hi Styfen,

Thank you very much for your clarification. It is clear.

Best,

Barbara