Optimizing and Satisficing Metrics

1

I have 2 losses: loss_a and loss_b

  • loss_a is a satisficing metrics. I just need loss_a < threshold.
  • loss_b is an optimizing metrics. I want it to be as low as possible.

Right now this what I am doing:

total_loss = alpha*loss_a + loss_b
total_loss.backward()

By adjusting alpha, I can adjust threshold for loss_a. Is there a better way to do this?

2

If you use lossa.clamp(min=threshold) it will only get a gradient when it’s above the threshold. That’s cool, but in reality you may have the other losses gradient pointing towards a direction of increasing lossa. How much this matters depends on your specific situation.

Best regards

Thomas

1 Like
3

This works pretty well for my application. Thank you