How to improve Classifica​tion/Regre​ssion generalization?

Hi,

I have a dataset 6x20000 as inputs and 100 outpus for a class problem.

The problem is that, the dataset is changing every second (not so much, but is different than the initially trained one).

Is there a way (a class/regress problem), to solve this kind of scenario, without counting with the "generalization" characteristic ?

Example:

Original trained input and output: 10 15 5 2 3 41 => 1

Original test input (not on the trained dataset) and output: 9 15 4 2 3 40 => 1.32

** "Changed" trained input: 9-11 14-16 4-6 1-3 2-4 2-5 40-42 => 1

Thank you.

André