version 1.9-alpha

The main new feature of version 1.9 is the availability of improved methods for computing the fit of models. A detailed analysis and presentation of the methods is given in Lartillot (2022).

More specifically:

• the old cross-validation approach has numerical inaccuracies and is not recommended anymore
• a new routine has been implemented for measuring the fit by leave-one-out cross-validation (loo-cv) and using the widely applicable (or Watanabe-Akaike) information criterion (wAIC).
• a method is also available for computing the marginal likelihood by sequential importance sampling; however, it is computationally intensive.

additional minor points:

• improved implementation and MCMC mixing for finite mixture models (-ncat option).
• bug corrected for restarting / reading finite poisson mixture models.
• bug corrected in old cross-validation under Poisson models.
• the number of components of truncated stick-breaking process is now set to 1000 -- and tunable on demand (backward compatibility implemented).

Lartillot, Nicolas. 2022. Identifying the best approximating model in Bayesian phylogenet-
ics: Bayes factors, cross-validation or wAIC? bioRxiv https://biorxiv.org/cgi/content/short/2022.04.22.489153v1

v1.8c

Very minor changes compared to v1.8b:

• chronometer of trace file compatible with more recent compiler versions
• compiling with -std=c++11
• use of std::numeric_limits::infinity() for dealing with 0 probability situations in logprob
• various warnings fixed
• one minor bug in readpb_mpi -allppred routine corrected (no change detected in behavior of the program, under fixed random seed)