Optimal parameters unstable in uncertainty model

Dear all:
I used the global optimizer cmaes to find optimal policy parameters in uncertainty shock model. I use a third order approximation and ‘pruning’ to generate the IRFs.
But the optimal parameters of each run are different, is cmaes not suitable to third order approximation?

The utility is utility=C^(1-sigma)/(1-sigma)-phi*N
totutility is totutility=utility+(1-beta)*totutility(+1)


If you are using simulated moments, then results may depend on the seed.