[Help Wanted] strange behavior with chains stuck in a non-optimal region of the tree topology space

[igoryanovich]

Describe the desired behavior.
I've tried to replicate a MrBayes analysis from doi:10.1163/22105832-bja10022 in RevBayes. The MB .nex file from the original paper is attached, as is my .rev recasting.

What is going wrong?
MB very quickly finds the optimal-likelihood segment of the tree topology space. RB behaves strangely: after the short burnin, each chain remains in the same likelihood region, see Tracer_RB_runs.png below.

RWTY comparison with MB chains shows that the best RB chains move in the same region as the MB chains, but the RB chains stuck at a lower likelihood. See rb_vs_mb.pdf attached.

Anything else you've tried?
I made sure to use a lot of SPR moves per generation; switched from simple MCMC to MC^3, trying deltaHeat=0.1 and 0.2, which did not help to get RB unstuck; with deltaHeat=0.5, hardly any changes between the chains were achieved (although I have not investigated that setting for a long time).

In MB, more tree moves are implemented than in RB (unless I'm missing something): in addition to NNI and SPR (they use "extended SPR" - I'm not sure if the regular SPR is meant, or something slightly different), MB also uses TBR and parsimony-biased SPR. I've tried to switch those moves off. The switching off of ParsSPR changes MB's behavior noticeably, see mb_all_variants.pdf below, compare the first four plots from chains w/o ParsSPR and the other four, with ParsSPR moves. However, this is not the only reason MB and RB are different, because even without ParsSPR, MrBayes finds the highest-likelihood segment which RB chains miss, cf. rb_vs_mb.pdf.

Screenshots

mb_all_variants.pdf
rb_vs_mb.pdf

Computer info
I tried this both on the Vienna VSC, and on my MacOS Laptop.

VSC:
RevBayes version (1.1.1)
Build from () on Tue Mar 28 19:44:44 CEST 2023

laptop:
RevBayes version (1.2.1)
Build from tags/v1.2.1 (04666d) on Mon Nov 7 15:45:19 UTC 2022

Additional context
Attaching analysis files - delete the added extension .txt

Original MB analysis file by Gunnink et al.: BantuMBgammaNY.nex.txt

My modification switching off ParsSPR and TBR moves: BantuMBgammaNY_wo_pSPR_TBR.nex.txt

My RevBayes recoding: SBa_MC3_for_issue.rev.txt

The Nexus file with the data - identical to the data part of the analysis file from Gunnink et al. SBantu-GunninkEtAl2023.nex.txt

[igoryanovich]

Something looks definitely wrong here: attaching the likelihood trace for four rather long parallel chains

(Before the chains start, there was a short burnin period, via mc3 burnin function.)

[davidcerny]

Perhaps @jsigao might know more?