NetRAX needs to return not only the best inferred network (in terms of BIC), but also best inferred network for 1 more reticulation
Closed this issue · 4 comments
I noticed that for some simulated datasets, NetRAX rejects (via BIC score criterion) the simulated true network in favor of less reticulations.
This often is due to our current simulations (things like ultrametric networks with a non-recoverable reticulation, or for example networks where two displayed trees only differ by a single branch length and this difference is very small).
Looks like if NetRAX settles for n reticulations in the end, we need to also write the best found (n+1)-reticulation network to file (together with its BIC and loglikelihood). This will help for evaluating the results.
In the NetRAX runs where I observed this (inferring less reticulations than in the dataset), the issue sometimes gets mitigated by increasing the size of the simulated MSA.