author question: auto-tuning evaluation features

[BubbaTough]

Howdy,

For those of you that successfully auto-tune your evaluation parameters(either by self play or move/score matching) how long/deep do you run per move? And are there certain parameters you find you need to spend more time on (or search deeper) during the tuning process?

I have been fooling around with a new auto-tuning approach for Hannibal, which I have started testing on some king safety features, and after some mediocre results on my first run started to vaguely remember some conclusion I had years ago that king safety was harder to auto-tune than many other features because it required more depth/time per move.

-Sam

[Ferdy]

Howdy,

For those of you that successfully auto-tune your evaluation parameters(either by self play or move/score matching) how long/deep do you run per move? And are there certain parameters you find you need to spend more time on (or search deeper) during the tuning process?

I have been fooling around with a new auto-tuning approach for Hannibal, which I have started testing on some king safety features, and after some mediocre results on my first run started to vaguely remember some conclusion I had years ago that king safety was harder to auto-tune than many other features because it required more depth/time per move.

-Sam

If it requires you more depth or time to tune your king-safety eval parameters, would it be because there are missing features that needs to be added, so that it is now aware even when search time is limited?