I would love to dig into the maximum expected score loss per move--what if we removed the situations where the player is +15 and then 'blundered' and ended up at +9 and still won? Like is there a way to filter out the moves where the score loss is practically meaningless?
I'm not using the evaluation directly, but the expected score, which I calculate from the evaluation. The expected score is then much more sensitive to evaluation changes in equal positions than completely winning ones.
In your example, the expected score would be above 99% in both cases, so the change wouldn't be significant, whereas going from 0.0 to +1 changes the expected score by about 20%.
I would love to dig into the maximum expected score loss per move--what if we removed the situations where the player is +15 and then 'blundered' and ended up at +9 and still won? Like is there a way to filter out the moves where the score loss is practically meaningless?
I'm not using the evaluation directly, but the expected score, which I calculate from the evaluation. The expected score is then much more sensitive to evaluation changes in equal positions than completely winning ones.
In your example, the expected score would be above 99% in both cases, so the change wouldn't be significant, whereas going from 0.0 to +1 changes the expected score by about 20%.