4. Experiments

Four small experiments were run both to (1) verify the general IVGA principle (Experiments 1 vs. 2), i.e, that it is useful to model independent variable groups separately, and (2) to study the performance of the presented ${\rm IVGA_{VQ}}$ algorithm in optimizing the model for the data and in finding independent variable groups (Experiments 1 vs. 3, Experiments 2 vs. 4).

In each experiment, the best one out of several trials is reported (for VQs this meant about 10000 trials, for ${\rm IVGA_{VQ}}$, considerably fewer). With each approach we used a roughly equal amount of computation time. In reporting results we report the number of codebook vector parameters--the number of the rest of the parameters is substantially smaller and constant in all experiments.