Supervised machine learning with out defined results?


I’m wondering if it’s possible to use machine learning to solve black box problems.
for example, if I have a black box that takes 6 inputs and outputs 6 value and i have a bunch of input-output sets generated from this black box, can I feed these sets into a machine learning algorithm so it can mimic the black box?
If so, what plugin can I use for this?

I’ve looked at lunchbox and crow, but their result have to be predefined, and te plugin just tells you which predefined result is most likely. But I dont want to predefine the possible results, cause lets say each of the 6 output is a number 1 to 180, I would have to define ten trillion possibilities…

(David Rutten) #2

If the network is dense enough you should be able to get it to at least replicate the training data. Whether it’ll give the right answer for new inputs is anyone’s guess. What percentage reliability you end up with depends on the continuity of the process inside the black box I presume.


If I understood correctly, you want to approximate an unknown black box only from input and output data of that black box. I guess that there is work in this regard, search about adversarial attacks, maybe guide you to what you’re looking for.

I think it’s possible if you have enough quantity and quality of input-output pairs, for example using genetic algorithms you could find the neural network that best simulates the original black box using supervised learning. But if you only have a biased subset of the data with which they have trained the black box (in case that the black box is a NN), you have little chance of getting a good approximation.

I hope I’m not helping you to do something illegal :see_no_evil: