1/ Some bugs are fun to discover!
A short #phdlife story.
A few weeks back, I found a bug in my dissertation model. It was a misspecification of a learning rule -- literally, I forgot a minus sign. This was a good find in that it made things correct but it was otherwise boring.
-
Show this thread
-
2/ However, I didn't rerun my main (big) simulations right away since I was working on other parts. The past few days, I've been doing so. Somewhat catastrophically, the runtime grew by a full order of magnitude! (On an already hard-to-parallelize model.)
1 reply 0 retweets 0 likesShow this thread -
3/ Without going too much into the weeds, I should say my dissertation agent-based model is sorta like a (very shallow) hierarchical model of Gibbs distributions.
1 reply 0 retweets 0 likesShow this thread -
4/ Energy goes down when agents episodically - Agree with Nature; - Agree with agents they trust (contingent upon experiential stereotypes) - Disagree they distrust.
1 reply 0 retweets 0 likesShow this thread -
💥 (wannabe) Ƀreaker of (the Bad) Loops 💫 Retweeted #TestAndTrace Smith 🐇
5/ There's a lot more details but I think this hints at the connection to things like negative partisanship and why fact checking doesn't work. (It's also formalizes a lot of things similar to
@Noahpinion's thread from yesterday.)https://twitter.com/Noahpinion/status/1039396322132938752 …💥 (wannabe) Ƀreaker of (the Bad) Loops 💫 added,
#TestAndTrace Smith 🐇Verified account @Noahpinion15/This is why you see Quillette, or the "Intellectual Dark Web" - people who have no allegiance to Trump or Fox News - focusing on and criticizing the idea of diversity. Oppositional thinking rules everything. A position reviled by my enemies must be one worth considering.Show this thread1 reply 0 retweets 1 likeShow this thread -
6/ Anyways, to sample over beliefs and stereotypes, I'm using stochastic acceptance, which has really nice properties for very dynamic distributions. https://jbn.github.io/fast_proportional_selection/ …
1 reply 0 retweets 0 likesShow this thread -
7/ So, why did my model slow down drastically? Because that subtle fix from two weeks ago induced drastic (and predicted!) skews in the agent distributions.
1 reply 0 retweets 0 likesShow this thread -
8/ The one that sunk performance was in the expectation distributions which are big (one per belief context) and determine activity selection. It went nearly power-law. The rejection rate skyrocketed!
1 reply 0 retweets 0 likesShow this thread
9/ I still don't have a decent fix. In fact, I was working on a paper for automatically balancing sampling (mixing strategies dynamically) in this scenario months ago before I shelved it for my dissertation work proper. But, it's kinda a neat a-ha!
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.