Maplesoft has gotten back to me: looks like some of the bugs we’ve hit are fixed in Maple 2019. So upgrading might be an easier path forward to a talk that works.

If you could test that the code I sent works, then we would already feel more comfortable giving that example in the talk.

Here are my plots using the artifact we submitted.

Please compare these results to the paper we submitted. There are three (minor?) discrepancies: 1. For NaiveBayesGibbs-Accuracy (not NaiveBayesGibbs-Likelihood), I had to change equal? to = in jags-accuracy and remove map sub1 in cleanup-jags. It seems that the JAGS output format has changed to become more in line with Hakaru and AugurV2 (i.e., 0.0 19.0 instead of 1 20)? Why? 2. In the GMM plots, STAN goes only until 10 seconds, even though in the paper we show STAN for 20 seconds. Why? 3. Hakaru and STAN seem slower in my GMM tests than in the paper…

1. I don’t know it’s been a long time since I looked at jags.
2. Oh I changed the code but I think I missed to check it in.
3. Hakaru became slower with the new version you uploaded, stan I don’t know.

1. By “the new version you uploaded”, do you mean changes to the Haskell/Maple side around last year"?

yes

Ok so I think we should (1) Fix 1 and 2 (2) Push changes to GitHub and tell artifact reviewers to pull (but what does @samth think?) (3) Regenerate all 6 plots on the actual machine (not VM) we say we use, and save the resulting logs

I can’t - I don’t personally have 2019 (yet). I’ll try to obtain a copy.