RBDO using hybrid GA for optimization - local and global cost plot

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

Hi All,

Thanks for this useful platform. Please appreciate your thoughts on this. I carried out a surrogate assisted RBDO using hybrid GA and had the following cost values for both the global and local cost

Global (2 iterations ) 0 (3.74) , 1 (3.162) , 2 (3.162)
Local (6 Iterations) 0 (3.91) , 1 (3.01), 2 (2.899) , 3(2.893) , 4 (2.892) ,5(2.892), 6(2.892)

Two cost plots are presented (global and local) and are different.

My question: Since the local optimization phase is meant to refine the outcome of the global, why is it that the cost value (local) at the initial iteration 3.91 is not the same as the convergence value of the global 3.162. Intuitively, I thought it should be a continuation from the convergence point of the global solution.

Appreciate any thoughts from the community, UQlab team and @moustapha @Team

Thanks always for the support and wonderful software. Your timely response will greatly be appreciated. Thanks All.

2

Hi @aokoro

The global optimizer runs a set of samples at each iteration (or generation). From memory I think that what you observe is the score for each generation. Depending on the algorithm, it may be the average of all responses, or their median. The refinement stage starts with the best sample from the last generation of the global optimizer. The corresponding response is therefore different from the average/median that is displayed/recorded by the global optimizer.

Please have a look at the fields .Results.History.GlobalOptim.X and .Results.History.LocalOptim.X to see what I mean with the number of samples per iteration in the global and local optimizers.

Hope this helps,
Moustapha

3

Thank you so much @moustapha for the clarification and your support always. Highly appreciated.