5.6.2 Combination of Local and Global Optimization Strategies

The promising results obtained from the example presented in the previous section seem to justify a combination of the two optimization strategies. As a global optimizer the very fast simulated re-annealing algorithm was used due to its better overall performance compared to the genetic algorithm. The same application as in the last section was used, but instead of only using a fixed number of evaluations, it was tried to (a) achieve the best target value obtained so far ($ \approx 8$) with the least possible number of evaluations and (b) to achieve a better target value with the given number of fixed evaluations.

The following algorithm was implemented to perform a combined optimization.

  1. Perform a configurable number of initial evaluations with the very fast simulated re-annealing algorithm.

  2. Start the gradient-based optimizer and use the best value of the very fast simulated re-annealing optimization so far.

  3. Wait for the gradient-based optimizer to exit and update the best target of the very fast simulated re-annealing algorithm if the target was improved.

  4. Terminate the optimization task if the truncation condition is fulfilled.

  5. Perform a configurable number of evaluations with the very fast simulated re-annealing algorithm. If the target was improved continue at step 2, otherwise at step 4.


Subsections

2003-03-27