Is the newton method necessary in loss based bwe v2?

[teck zhou]

Dear all,

I can't figure out why we need newton method to search the best candidate. My derivation as below:

Is there any problem with my derivation? Or have I misunderstood something?

[il...@google.com]

I'm not sure about the Newton method, but in your calculations there's a mistake at the last step. It should be ratio of two sums, not the sum of ratios:

p = sum(n_i) / sum(n_i+m_i)

[Ilya Nikolaevskiy]

+Fanny Linderborg , who's authored the code. I think the Newton method is used since the minimized function isn't as simple as you've stated.

It also takes the current bandwidth estimate and actual sent rate into account. So maybe such a simple formula isn't even possible to derive there.

--

Best regards,

Ilya

[Jason Gu]

As I understand, first, inherent_loss is not loss_probability; second, the purpose is not to calculate the loss_probability, but the loss_limited_bandwidth.

[Jason Gu]

loss_probability = inherent_loss + (sending_rate - loss_limited_bandwidth) / sending_rate;

replace p with loss_probability in your function and do d((ln(L(p)))/d(inherent_loss)

[teck zhou]

I think you are right, the observation loss may comprise inherent loss and link capacity limit loss.  In other words, both the loss limited bandwidth and the inherent loss are unknown, and we need to estimate the most probable combination of loss limited bandwidth and inherent loss based on the current observation data. Is my understanding correct?

在2024年5月25日星期六 UTC+8 09:18:58<Jason Gu> 写道：

[Jason Gu]

Yes I agree with you. There should be two variables:  inherent_loss  and loss_limited_bandwidth. So partial derivative should be used. I don't know why the algorithm in loss_based_bwe_v2 of webrtc only use inherent_loss.