1
$\begingroup$

I would like to do production optimization with machine learning and/or optimization problem.

My goal is not to find minimizing loss in loss function only to give the best y value. My ultimate goal is to find optimal values of the best feature combination. In other words, 'I would like to find the best x values (feature values) that yields the best y (minimizing loss). For example, 30 (temp), 100 (Pressure), 50 (weight) gave the best product y. The actual best set of x feature values.

Supervised machine learning techniques do not provide the best-predicted x values to me. but I would like to find the x. Therefore, my question is as below:

  1. What is the field for finding optimal x values for optimal y values called? Is continuous and discrete optimizations are ones finding the x values for y?

  2. I found many studies but I am not sure if it is really about finding the x values (the optimal values of the best combination of features). Can someone please tell me what field I must study for this?

  3. What kind of techniques should I use to find what I want?

Improve this question
$\endgroup$
1
  • $\begingroup$ could you please share the approach you followed ? specifically im interested in the objective function. $\endgroup$
    – Siddhant Tandon
    Dec 18, 2020 at 8:30

1 Answer 1

Reset to default
1
$\begingroup$

Welcome to DS.SE!

As you already mentioned, machine learning is not about finding the best values of features regarding a function. Also, loss functions usually measure the errors of predictions and not the quality of the feature values (e.g. " the best product y").

What you are looking for is called optimization or search algorithms but not learning or prediction. If the search space (i.e. the combination of all possible feature values) is not that large, you can just perform an exhaustive search (trying all combination). If the search space is large, then smarter approaches might be useful. These techniques need a function (usually called objective function) to optimize (finding the combination of input values that results in a min/max outcome).

In general, there are two approaches: mathematical modeling (i.e. linear programming) and metaheuristics (e.g., evolutionary algorithms and swarm intelligence). Mathematical models can be fast and guarantee to find the optimum when the objective function is well-defined and the underlying assumption are met (you may need to simplify your constraints sometimes). If these conditions are not met, then metaheuristics usually can find a set of approximate solutions.

Improve this answer
$\endgroup$
4
  • $\begingroup$ Thank you for your great answer. I should definitely look at the optimization algorithms! Cheers! $\endgroup$
    – junmouse
    Jun 11, 2019 at 9:00
  • $\begingroup$ I have a question again. As Reinforcement Learning is quite similar to search problems, can RL be a method to my question as well in a way? Thank you for your feedback in advance. @Borhan $\endgroup$
    – junmouse
    Jun 11, 2019 at 9:39
  • $\begingroup$ RL is a specific type of online problem: you have an agent (solver) that acts in a dynamic and mostly unknown environment. It tries to maximize long term reward (or minimize regret). At each time, the agent decides to explore the environment to learn more about it or take the action that knows a good one from experience (more or less, similar to how humans learn from their mistakes/tries). So, if your problem dynamically change and you have to adapt to changes rapidly, RL might be helpful. However, from the example you provided, I guess optimization techniques do the job for you. $\endgroup$
    – Borhan Kazimipour
    Jun 11, 2019 at 21:27
  • $\begingroup$ Thank you very much for your help. This boosts up. Many thanks! $\endgroup$
    – junmouse
    Jun 11, 2019 at 23:49

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.