If this is a duplicate, I apologize. I'm not really sure what to even search for to try and find a duplicate/answer!

We are working on a system for providing musical feedback to change the 'mood' of a subject. There are a number of parameters of the music that we can manipulate.

At the same time, we measure the physiology of the subject. I'm using a dynamic Bayesian network to reasonably accurately determine the level of frustration of the subject.

I'm looking for a way to, based on the level of frustration, tell the music generation mechanism "this set of parameters is effective", or not. The music parameters would then adjust automatically and iteratively, as we continue to make judgements of the subject's level of frustration.

I'd appreciate any pointers I can get! Please let me know if I can provide any clarification!

  • 1
    $\begingroup$ In a sense you want to update the network, ie online learning. $\endgroup$ – Adam Bittlingmayer Oct 12 '15 at 16:57

This can be thought of as an active learning problem. You want feedback from the learning process to influence what kind of input examples you next train on. Active learning is still a difficult problem in most settings.

As @cohoz points out, hill climbing is an intuitive option in this case, but only if you always want to increase or decrease your measure of frustration. If instead the goal is to explore frustration as a function of as much of the music parameter feature space as possible, you will want to read about active learning and uncertainty.

|improve this answer|||||
  • $\begingroup$ Could you provide some references to get started on Active Learning? $\endgroup$ – A. G. Feb 2 '18 at 23:56

One basic approach would be a hill-climbing algorithm to search for local optima.

This will be more or less effective depending on the complexity of the space (e.g. convexity: in theory will this find the local optima, but also more practical concerns around the size/dimensionality of the space and easy of specifying neighboring points and the accuracy/precision of the frustration measurement).

This also answers a slightly different question: the acceptance criteria "this set of parameters is effective" seems linked to some threshold of "frustration" that you would define as opposed to simply a process to "minimize the frustration function with respect to these parameters", although one can view the latter as the way to make the iterative adjustments to search for the former.

|improve this answer|||||

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.