1

First off, I am new to machine learning, so these questions may be trivial.

Basically I am trying to tune an object with numeric knobs and numeric outputs. By doing a brute force tuning (permutations), there is a solution to find ideal values of the outputs, but it takes time. I'm trying to take a stab at using ML to at least shorten the process of tuning.

object to tune

I have a dataset for a large number of good units that were tuned successfully, at the various number of attempts though.

For example: passing requirement: X = near 10, Y = near 5, Z = near 4

Object 1:

Try 1 => A = 1, B = 2, C = 3 ; X = 1, Y = 0, Z = 1  => not good
Try 2 => A = 1, B = 1, C = 1 ; X = 10, Y = 5, Z = 4 => good enough

Object 2:

Try 1 => A = 1.4, B = 2.6, C = 3.8 ; X = 10, Y = 5, Z = 3.9 => lucky!!!

Object 3:

...
Try 10 => A = 1.4, B = 2.6, C = 3.8 ; X = 10, Y = 5, Z = 3.9 => took a while!!!

I'm wondering how do I prepare the data for training and testing with this type of problem, as each object had a variable number of tries before it got successfully tuned. Should I just take the last successful combination for each object, and maintain the same columns (A, B, C, X, Y, Z). Or take them all (multiple rows per object)?

Or, for each object record, append another set of columns, such that per object, I have only one row. For example (A1,B1,C1,X1,Y1,Z1, A2,B2,C2,X2,Y2,Z2,... An,Bn,Cn,Xn,Yn,Zn)

As for the algorithm choice, I can only articulate that this isn't a classification problem (or a binary outcome). Something like regression, or decision tree, or hill climbing if there's such a thing?

Darshan Jain
  • 205
  • 1
  • 10
alpinescrambler
  • 119
  • 2
  • There are lots of possibilities depending on expected shape of function. If it varies smoothly enough then some form of hill climbing would be most efficient. But if the function shape is highly complex, then there might be no better solution than fine-grained brute-force search, or even just random guesses. I think there is a duplicate question for complete "black box" tuning. – Neil Slater Mar 30 '17 at 16:40
  • However, before marking as duplicate (which might answer your question), I'd like to know more context. Is this part of a course assignment? If so, what are the topics - because you'd usually be expected to use techniques that you had studied? If not, please explain what else, if anything, you have been told about this box or the function inside - is it a real physical box, or some kind of simulation? Have you been told any characteristics of the function at all, such as smoothness? – Neil Slater Mar 30 '17 at 16:44
  • Not a part of a course assignment. Looking at my data, there is somewhat a one-to-one correlation in terms of one particular input to an output. For example, decreasing A, would directly increase X, but it also "changes" Y, & Z. This is a real physical box. – alpinescrambler Mar 30 '17 at 16:50
  • What is the goal of learning? It is not clear. E.g. is it to minimise number of measurements taken on new boxes to tune them? Is it to find a tuning setting that should work on any box? Box 1 seems different to boxes 2 and 3, but boxes 2 and 3 give identical results on the tuned position - is that correct (or you just using made-up examples)? – Neil Slater Mar 30 '17 at 19:22
  • I have 2 problems. 1.) How to clean the data in order for the training process to consume. 2.) What algorithm . These are just made up examples. – alpinescrambler Mar 30 '17 at 23:39
  • Picking an algorithm will probably not solve an issue that you are not able to articulate. I have a problem 0 - what exactly is it you are trying to do? It still is not clear to me, but perhaps someone else will be able assist from what you have said so far. However, perhaps this question (which I have answered) is related enough and might help: http://datascience.stackexchange.com/questions/8043/optimizing-parameters-for-a-closed-black-box-system – Neil Slater Mar 31 '17 at 10:19

1 Answers1

1

I don't know that I'd call this a machine learning problem per se...though others may disagree, this sounds like an optimization problem.

Your description is vague, but let's pretend it's a guitar that you're trying to tune and you have 6 strings. The trick is to define an objective function that quantifies, with a single number, how 'good' the tuning is. (The program can independently manipulate each string's tension). Perhaps it's some deviance in power from an ideal FFT, and you sum the deviance from each string.

Optimization routines such as simplex, sub plex, Nelder-Mead, GA, etc., are ways of manipulating the tension in 'smart' ways that obviate the need for a brute force (or 'grid search') solution. So long as you can define an objective function, you can start trying it out, though the local max/min problem will need to be dealt with.

I'd look into some tutorials on minimization algorithms

HEITZ
  • 911
  • 4
  • 7