Number of starting networks in NetRAX

[lutteropp]

We agreed earlier that all random starting networks should be random trees.

5 random starting networks are not enough, even if the "true" network is a tree. RAxML-NG uses by default 10 parsimony + 10 random starting trees. I tried NetRAX (with only 5 random start networks) on a small (10 taxa) tree dataset and it got stuck in a local optimum.

I suggest copying the default from RAxML-NG and also using 10 random starting trees + 10 random parsimony trees in NetRAX.
For later, we could think about replacing the parsimony trees by parsimony networks.

[lutteropp]

How about these two experimental setups:

A: Start NetRAX from 10 random + 10 parsimony trees
B: Start NetRAX from only the best RAxML-tree

[lutteropp]

I implemented the two setups I suggested above. Let's see how it behaves in our experiments now! :)

[stamatak]

yes that is what is was about to suggest, we need to explore broadly first to obtain a feeling for who this behaves

…

On 25.11.20 18:35, Sarah Lutteropp wrote: I implemented the two setups I suggested above. Let's see how it behaves in our experiments now! :) — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#7 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAGXB6XTFNCXJ7YCQ76NSCTSRUW3TANCNFSM4UCROIFQ>.

-- Alexandros (Alexis) Stamatakis Research Group Leader, Heidelberg Institute for Theoretical Studies Full Professor, Dept. of Informatics, Karlsruhe Institute of Technology www.exelixis-lab.org

[lutteropp]

Turns out 10 random + 10 parsimony trees are a bit much... especially since the current NetRAX version is only single-threaded and not optimized for runtime performance yet (it wastes much time in branch-length optimization and in trying and rejecting arc insertion moves). Trying out what happens if I reduce them by a lot.

[stamatak]

just use brute force and the cluster for your experiments

…

On 30.11.20 02:02, Sarah Lutteropp wrote: Turns out 10 random + 10 parsimony trees are a bit much... especially since the current NetRAX version is only single-threaded and not optimized for runtime performance yet (it wastes much time in branch-length optimization and in trying and rejecting arc insertion moves). Trying out what happens if I reduce them by a lot. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#7 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAGXB6S3MUF2U5DPU6T4VZLSSLOJRANCNFSM4UCROIFQ>.

-- Alexandros (Alexis) Stamatakis Research Group Leader, Heidelberg Institute for Theoretical Studies Full Professor, Dept. of Informatics, Karlsruhe Institute of Technology www.exelixis-lab.org

[lutteropp]

If we don't care aabout runtime at all: Should I also switch to the slower network search algorithm? So far, I accepted the first move that improved the BIC score. But I already observed that I get a bit better results if I evaluate the entire 1-move-neighborhood and then accept the move that lead to the largest improvement in BIC score...

[stamatak]

we don't care about run time for the time being, however applying several moves earlier and quicker may also lead to the desired result, ... to be explored

…

On 30.11.20 11:13, Sarah Lutteropp wrote: If we don't care aabout runtime at all: Should I also switch to the slower network search algorithm? So far, I accepted the first moce that improved the BIC score. But I already observed that I get a bit better results if I evaluate the entire 1-move-neighborhood and then accept the move that lead to the largest improvement in BIC score... — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#7 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAGXB6RQ5VOYC5LZEYATNBLSSNO3RANCNFSM4UCROIFQ>.

-- Alexandros (Alexis) Stamatakis Research Group Leader, Heidelberg Institute for Theoretical Studies Full Professor, Dept. of Informatics, Karlsruhe Institute of Technology www.exelixis-lab.org

[lutteropp]

With the new wavesearch algorithm, I do not see any advantage in starting from multiple starting trees vs. starting from best raxml-ng tree.

[stamatak]

see discussion on slack, starting from several RAxML-NG optimized trees is something we should test

…

On 17.01.21 02:58, Sarah Lutteropp wrote: With the new wavesearch algorithm, I do not see any advantage in starting from multiple starting trees vs. starting from best raxml-ng tree. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#7 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAGXB6VHOECRMYDU6F4EENTS2IY4HANCNFSM4UCROIFQ>.

-- Alexandros (Alexis) Stamatakis Research Group Leader, Heidelberg Institute for Theoretical Studies Full Professor, Dept. of Informatics, Karlsruhe Institute of Technology www.exelixis-lab.org