Question about cluster centers RBF basic

[Andrs]

During the training phase, the w parameters (w0,w1,w2...wn)are calculated using the training data (as well as the corresponding cluster centers and distance to the training_x points).
During the test phase, new fresh data is generated. In order to get the E_out for RBF basic, we use the w-parameters that were "learned" during the training phase.
During the test phase, we have also generate a new "fi-transform" that contains the distance between the cluster centers and the new x_test points. Do we calculate new cluster centers with the test data (we have all info to calculate new cluster centers with the test data)? Or do we re-use the cluster centers (muk) calculated during the training???? I tried both and I can see some very definite diffs in the results...but I am wondering if the cluster centers defined with training data are representative in the test....

[yaser]

Originally Posted by Andrs

Do we calculate new cluster centers with the test data? Or do we re-use the cluster centers (muk) calculated during the training???? The cluster centers defined during training may not be representative for the test data...or?

You use the cluster centers calculated during training. All parameters are estimated in sample, and whether or not they work on the test data is the usual question of generalization.

[JohnH]

The cluster centers and weights ( and , respectively), along with the RBF, comprise the learned function. Once learning is complete, and are fixed. is part of a measure of the the quality of the learned function: it supports claims that the learned function is PAC (Probably Approximately Correct, in other words, it is a good generalization).

[Andrs]

Thanks, the answer matches also more reasonable results!