Reinforcement learning, mostly. The parts I keep circling back to are multi-agent learning, environment design, and the question of when an agent should stop gathering information and commit.
Mouhssine Rifaki
Karma: 9
Reinforcement learning, mostly. The parts I keep circling back to are multi-agent learning, environment design, and the question of when an agent should stop gathering information and commit.
Thanks for this, this is a concrete post. So the selection step itself is treated pretty informally. Parents either set hard thresholds or build a weighted spreadsheet and either way we’re collapsing a pretty rich posterior into a single number per embryo, the reports already carry more information than that approach uses. Before touching weights at all, we can drop embryos that are worse on every trait the couple cares about. If an embryo is strictly dominated by another, there’s no preference structure under which it’s the right pick so we can eliminate it geometrically and remove a chunk of “did we pick the wrong weights” regret for free
and when parents do have to weigh the remaining embryos, the usual ‘how many iq points equal one percent of schizophrenia risk’ framing treats each trait as independent and flattens the variance, I find the fritz story is a nice illustration of the alternative mattering in practice: so the t1d posterior is bimodal because of the protective variant not just low on average and an expected-value comparison would have obscured how onesided that call actually was
Do any of the companies expose the variance on individual predictions or only the point estimates? 9% ± 0.5% is a different decision than 9% ± 4% when the alternative embryo is at 7%, and i’d guess most parents would rather take a lower mean embryo with tight variance than a higher-mean one with wide variance if the report actually let them see that, if the variance is in the underlying predictor but not in the report, that seems like a cheap product fix imo