the histogram of snarks by waking-time as a mixture of gaussians
but was unable to make much progress; my guess is that either
there are 2+ classes with very high variance, polluting everything, or maybe some snarks have a different distribution with a longer right tail than normal.
I did note that
Looks like we don’t hunt Crumbling and are strongly biased against hunting Blunt, so maybe I’ll just keep to conventional wisdom there. Of the remaining, very close to exactly 3% aren’t hunted, with no discernible correlations? Maybe that’s just the baseline “don’t hunt this” chance.
So I fell back to
First remove all Crumbling and Blunt, then Just pretend you can treat every variable as independent, then pretend you can treat every pair of variables as independent after accounting for first-order effects. Grab log-likelihoods and smush them all together.
Which led me to
Not sure how many to hunt, but in this order: [‘V’, ‘Y’, ‘G’, ‘P’, ‘Q’, ‘L’, ‘H’, ‘W’, ‘C’, ‘M’, ‘B’, ‘N’, ‘J’, ‘R’, ‘D’, ‘X’, ‘K’, ‘F’]
The two suggested criteria would lead to either [‘V’, ‘Y’, ‘G’, ‘P’, ‘Q’, ‘L’] (maximize my survival), or [‘V’, ‘Y’, ‘G’, ‘P’, ‘Q’, ‘L’, ‘H’, ‘W’, ‘C’, ‘M’, ‘B’, ‘N’, ‘J’, ‘R’, ‘D’] (maximize snark count EV)
Had I submitted in time, I probably would have chosen to stop at either H, W, or C, because my estimated boojum probabilities have an inflection point there. If I were trying to “beat” everyone else, I’d have stopped at C; if I were ignoring everyone else, I’d have stopped at H, which, interestingly, is only one more snark than the bare minimum.
I tried for a while to identify
the histogram of snarks by waking-time as a mixture of gaussians
but was unable to make much progress; my guess is that either
there are 2+ classes with very high variance, polluting everything, or maybe some snarks have a different distribution with a longer right tail than normal.
I did note that
Looks like we don’t hunt Crumbling and are strongly biased against hunting Blunt, so maybe I’ll just keep to conventional wisdom there. Of the remaining, very close to exactly 3% aren’t hunted, with no discernible correlations? Maybe that’s just the baseline “don’t hunt this” chance.
So I fell back to
First remove all Crumbling and Blunt, then Just pretend you can treat every variable as independent, then pretend you can treat every pair of variables as independent after accounting for first-order effects. Grab log-likelihoods and smush them all together.
Which led me to
Not sure how many to hunt, but in this order:
[‘V’, ‘Y’, ‘G’, ‘P’, ‘Q’, ‘L’, ‘H’, ‘W’, ‘C’, ‘M’, ‘B’, ‘N’, ‘J’, ‘R’, ‘D’, ‘X’, ‘K’, ‘F’]
The two suggested criteria would lead to either
[‘V’, ‘Y’, ‘G’, ‘P’, ‘Q’, ‘L’] (maximize my survival), or
[‘V’, ‘Y’, ‘G’, ‘P’, ‘Q’, ‘L’, ‘H’, ‘W’, ‘C’, ‘M’, ‘B’, ‘N’, ‘J’, ‘R’, ‘D’] (maximize snark count EV)
Had I submitted in time, I probably would have chosen to stop at either H, W, or C, because my estimated boojum probabilities have an inflection point there. If I were trying to “beat” everyone else, I’d have stopped at C; if I were ignoring everyone else, I’d have stopped at H, which, interestingly, is only one more snark than the bare minimum.