Full disclosure, I also posted this on Stack Overflow I have put a more theory based bent towards the question itself here

I have a simple model in pytorch based on the quickstart except instead of a pre-made data-set I am trying to find parameters for a mathematical function to fit a curve, I am then trying to feed back into the network the "goodness of fit" through the loss for the curve. My output is (6 parameters) for the function.

I gave as input:

  • The independent input at N inputs (lets call it x), this does not change
  • The dependent output at N outputs (lets call it y(x)), this does not change
  • The output of the mathematical function f for each point of x, and parameters p (lets call it f(x, p) where p is the same for all x)
  • The last set of guessed parameters p

Now I have a simple cross entropy loss function, and stochastic gradient descent for back-propagation in a simple feed forward network, for which the output is a new set of 6 parameters for f, p', I take p' and x and put it into f and back-propagate based on y(x) and f(x, p'), not p, loss is indirectly based on the output.

Behavior I am observing is that the output does not change very much, really just the lower digits (also should it be outside the bounds [-1, 1]), neither has the loss. I have made attempts to divide it out by the absolute value of the absolute maximum value of the output when feeding it back into the network but no luck

Questions: A: Am I doing anything fundamentally wrong? Can I feed the previous results back into a network and expect a change? Can I sort of feed it a "score" to tell it how well its doing? What kind of model may be able to do this, if not feed forward? B: Could this be a normalization problem? C: How might I diagnose this problem?

I'm trying to do a sort of "live" learning/generation or "active" learning perhaps.


  • $\begingroup$ Looking at genetic algorithms, or genetic + NN, also possibly baysian optimization, and NEAT $\endgroup$
    – cgbsu
    Commented Nov 13, 2023 at 3:34


You must log in to answer this question.