What I did today:
1. Realized that the worst case when you condition on something is arbitrarily bad (which I think I figured out a while ago, so feel stupid for trying this again)
2. Tried to salvage this by using the fact that the polytope is something that you have control over
3. Have what I think is a decent approach to this, but it may not matter because it seems like another issue is the relaxation you get from considering a valid abstraction of your likelihood instead of the likelihood itself — it can lead to really extreme assignments of probabilities.
4. Thought a bit about how to fix this, considering just trying to learn the correct likelihood from data, although it seems possible that there’s something better to be done.
Also, I explained kernel methods and concentration inequalities to Qiaochu and met with Percy for a couple hours.