Reason for dropping the previous tree

[tom94]

Reason for dropping the previous tree

[Tom94]

Hi Shilin,

no problem at all!

Absolutely: dropping the previous tree is creating a certain level of redundancy. However, due to the power-of-2 sample count scheme, only at most half of all samples are thrown away. So the maximum amount of information you can save by re-using old trees is 2x.

One possible option would be an exponential moving average (e.g. as used by Dahm and Keller [2018]), but exponential moving averages also have downsides. For one, they always emphasize the latest samples, exaggerating their variance. Plus, from my experience, it was difficult to accurately shape the re-use of previous information vs. the use of new information.

In any case, it's an interesting open problem!

[ShilinC]

Thanks! Very appreciated!