k = 8 past bandwidth and Next chunk sizes

[ali1hammoud]

I read and quote:

4.4 Implementation To generate ABR algorithms, Pensieve passes k = 8 past bandwidth measurements to a 1D convolution layer (CNN) with 128 filters, each of size 4 with stride 1. Next chunk sizes are passed to another 1D-CNN with the same shape

How did you choose k to be 8? why isn't another value? how do we determine this value or according to what? can you mention it in the code "k = 8 past bandwidth and Next chunk sizes" ?? Thank you.

[hongzimao]

Good question. In development, we tested a few values - for k < 8 there might be a performance degradation, for k >=8 the performance doesn't change much - in essence, we just used the minimum number past observations that worked. Hope this helps.

[ali1hammoud]

Yes, thank you. That was helpful.