Thank you very much @styfen.schaer
makes sense, thank you very much for your help:)
I have one more question. After reading this Acceptable LOO Error about acceptable LOO error I am wondering this.
I am performing LARS-based PCE. In each iteration, a highly correlated polynomial is added to the design matrix, and I compute the minimum modified LOO error (Eq. 1.22 in the UQ manual) for a nonlinear QoI with 10 stochastic inputs. The LOO error is not normalized, and I obtain the following results for a maximum chaos order of 2.
for:
100 Training Samples: modified LOO Error is 0.0022469228
50 Training Samples: modified LOO Error is 0.0145296080
25 Training Samples: modified LOO Error is 0.0200713039 (here I noticed >=10% errors when computing standard deviation in comparison with the Monte Carlo results)
So, knowing that this is a problem-dependent decision, is there a general empirical rule that indicates which results are trustful? Is the modified loo error a capable indicator to say that?
I want also ask about the max order of chaos. I have chosen 2 for this case. In LARS with a small training sample like here is there also a rule of thumb for that?