Multi-arm bandits work well in situation where you have choices and you are not sure which one will maximize your well being. You can use the algorithm for some real life situations. As an example, learning can be a good field:

If a kid is learning carpentry and he is bad at it, the algorithm will tell him/her that he/she probably should need to move on. If he/she is good at it, the algorithm will tell him/her to continue to learn that field.

Dating is a also a good field:

You're a man on your putting a lot of 'effort' in pursuing a lady. However, your efforts are definitely unwelcomed. The algorithm should "slightly" (or strongly) nudge you to move on.

What others real-life situation can we use the multi-arm bandit algorithm for?

PS: If the question is too broad, please leave a comment. If there is a consensus, I'll remove my question.

  • 3
    $\begingroup$ Given that there are 3 upvoted answers (so far), I don't think this is too broad to be answerable. $\endgroup$ – gung Aug 18 '16 at 18:36
  • $\begingroup$ @gung I've got more upvotes and yet, they are not reflected on my score. How come? $\endgroup$ – Andy K Aug 18 '16 at 19:33
  • 5
    $\begingroup$ That's because this thread is community wiki (CW), @AndyK. When a thread is CW, people don't get reputation from upvotes (or lose it from downvotes). You would earn badges as normal, though. Questions like this that solicit lists of stuff & where there isn't a single, clear 'correct' answer are supposed to be off topic on SE sites. Our compromise (I believe other sites do this too) is to allow such questions on a case by case basis, but to make them CW. $\endgroup$ – gung Aug 18 '16 at 19:38
  • $\begingroup$ fair enough @gung $\endgroup$ – Andy K Aug 19 '16 at 8:42
  • 1
    $\begingroup$ college admissions. Selection of metrics for selection of recipients for donated organs. $\endgroup$ – EngrStudent Aug 20 '16 at 20:04

When you play the original Pokemon games (Red or Blue and Yellow) and you get to Celadon city, the Team rocket slot machines have different odds. Multi-Arm Bandit right there if you want to optimize getting that Porygon really fast.

In all seriousness, people talk about the problem with choosing tuning variables in machine learning. Especially if you have a lot of of variables, exploration vs exploitation gets talked about. See like Spearmint or even the new paper in this topic that uses a super simple algorithm to choose tuning parameters (and way outperforms other tuning variable techniques)


They can be used in a biomedical treatment / research design setting. For example, I believe q-learning algorithms are used in Sequential, Multiple Assignment, Randomized Trial (SMART trials). Loosely, the idea is that the treatment regime adapts optimally to the progress the patient is making. It is clear how this might be best for an individual patient, but it can also be more efficient in randomized clinical trials.

  • $\begingroup$ Thanks @gung. I did not know about that algorithm. I will have a read at it $\endgroup$ – Andy K Aug 18 '16 at 15:31

They are used in A/B testing of online advertising, where different ads are displayed to different users and based on the outcomes decisions are made about what ads to show in the future. This is described in nice paper by Google researcher Steven L. Scott.


I asked the same question on Quora

Here's the answer

  • Allocation of funding for different departments of an organization

  • Picking best performing athletes out of a group of students given limited time and an arbitrary selection threshold

  • Maximizing website earnings while simultaneously testing new features (in lieu of A/B testing) You can use them anytime you need to optimize results when you don't have enough data to create a rigorous statistical model.


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.