Why the order of input variables has impact on the performance of Multikan？

[chen-erqi]

In my case, the mape results among different orders of input variables are huge different.
Such as:

col_array = np.array([0, 5, 1, 4, 2, 3]) # 8.69%

col_array = np.array([0, 1, 2, 3, 4, 5]) # 19.67%

col_array = np.array([5, 0, 1, 2, 3, 4]) # 14.50%

Could anyone please to discuss this interesting phenomenon?
Thanks a lot!!

[chen-erqi]

I can not see this phenomenon in the KAN without multi operation.

[chen-erqi]

The width of multiKan is [6, [0, 1]].
When I add a KANLayer before input layer, this phenomenon has weaken but not disappeared.
Such as:
col_array = np.array([0, 5, 1, 4, 2, 3]) # 3.34%
col_array = np.array([0, 1, 2, 3, 4, 5]) # 3.13%
col_array = np.array([5, 0, 1, 2, 3, 4]) # 3.18%

[KindXiaoming]

maybe I didn't get the point, but I think change the order of input variables = change initialization, which makes results different.