AAD adaptative power

[pruzhinskaya]

About the possibility to increase the impact of expert feedback on the AAD model.

From Shubhomoy:
https://github.com/shubhomoydas/ad_examples/blob/master/ad_examples/aad/aad_loss.py
The parameter prior_influence in the two functions (aad_loss_linear and aad_loss_gradient_linear) might help.

This parameter needs to be passed through aad_weight_update (https://github.com/shubhomoydas/ad_examples/blob/78b01e9c9502523c5341243e1a8dca6befcefbc3/ad_examples/aad/aad_base.py#L260)

[matwey]

Do you have a plan how to test it? I cannot figure out which prior influence value is big enough. By default it is 1.0.

[pruzhinskaya]

Yes, we definitely have to test it using a smaller dataset.

[matwey]

I don't see any significant difference when running with 1.0, 10.0, and 100.0.
I use Y for all choices. Is it correct?

[pruzhinskaya]

If you say YES the model should not changed, it is changing when you say NO

[matwey]

I suggest to discuss at the evening, that we all understand how this is going to work in the same way.