I don't understand the number of schemes analysed for the greedy algorithm

[iqtree]

I don't understand the number of schemes analysed for the greedy algorithm

[roblanf]

OK, there are two issues here.

FIrst, —merge greedy isn’t working to force it to do the greedy algorithm. The fix is to make sure that when the default -rcluster-max is specified, it doesn’t overwrite the —merge option. The upshot of this issue is that the algorithm here was rcluster, not greedy.

Second, we now understand why the number of schemes wasn’t right, even for the rcluster algorithm. We have a cool new feature for the rcluster algorithm which we forgot about! In the original in partition finder, we take the e.g. 10% of closest pairs based on absolute difference. THe issue with that is that we’ll choose lower-rate pairs preferentially, because they have the lower difference. So we’ve added a feature where we take the union of two sets:

1. The rcluster-max closest in absolute rate difference
2. The rcluster-max closest in log(rate) difference

The upshot is that the number of schemes in every step is between rcluster-max and 2*rcluster_max

[roblanf]

@bqminh can you post here when --merge greedy is working. Details in comment above.

[thomaskf]

Hi Rob,

Thanks for reporting the issue. It was due to an upper bound introduced for the number of candidates when computing the best-fit partition scheme.

In version 2.2.2.8 (https://github.com/iqtree/iqtree2/releases/tag/v2.2.2.8), the issue has been fixed and the greedy algorithm for computing the best-fit partitioning scheme (i.e. option: --merge greedy) should work as expected.

Thomas