2
$\begingroup$

Let's say that I want to create an optimization algorithm, which is supposed to find an optimum value for a given objective function. Creating an optimization algorithm to explore through the search space can be quite challenging.

My question is: can machine learning be used to automatically create optimization algorithms? Is there any source to look at for this?

$\endgroup$
1
  • $\begingroup$ if you can explain the challenges that you know, or even break it down into parts, it will give useful hint of some sort of structure for the answer $\endgroup$
    – Sanyou
    Nov 29 '21 at 16:09
3
$\begingroup$

Machine learning has been used to automatically learn new optimization/learning algorithms. This task is often known as meta-learning, i.e. you learn to learn, in this case, an optimization algorithm, but note that meta-learning does not just refer to learning optimization algorithms (see this blog post).

The blog post Learning to Optimize with Reinforcement Learning (2017) is a good introduction to the topic and focuses on the approach proposed in this paper Learning to Optimize (2016), which uses reinforcement learning to solve this problem: more specifically, they learn a policy (in practice, represented as a neural network) that represents the learned optimization algorithm.

There are other related approaches: for example, you may be interested in the paper Learning to learn by gradient descent by gradient descent (2016, NeurIPS).

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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