Hi Sergei,

On optimising models using Yang's M1a (neutral) and M2a (selection) omega distributions, I frequently get higher likelihoods for the constrained model. This seems to point to an optimisation problem, which is perhaps not very surprising as I am using an unusually large number of parameters (my model is probably overparameterised). The problem does not appear with more usual models.

How can I play around with the optimisation criteria? For instance, is there a likelihood threshold that determines at which point the optimisation stops? Also, is the algorithm completely deterministic or is it worth running it more than once to see if it always converges to the same answer?

regards
Konrad

Dear Konrad,


wrote on Nov 3^rd, 2005 at 5:46am:

On optimising models using Yang's M1a (neutral) and M2a (selection) omega distributions, I frequently get higher likelihoods for the constrained model. This seems to point to an optimisation problem, which is perhaps not very surprising as I am using an unusually large number of parameters (my model is probably overparameterised). The problem does not appear with more usual models.

This seems odd. There should not be inordinately many parameters in these two models; perhaps some parameters are being declared as local, inadvertently? You should try to list all model parameters (e.g. using the -p command line option and choosing View Likelihood -> Show Parameter list at the end of the run if using a command line version; or using the Object Inspector for the appropriate likelihood function to bring up the parameter table in the GUI) and see if any of them are extraneous.


Quote:

How can I play around with the optimisation criteria? For instance, is there a likelihood threshold that determines at which point the optimisation stops? Also, is the algorithm completely deterministic or is it worth running it more than once to see if it always converges to the same answer?

Default optimization algorithm is determinstic (given a starting point).
Some of the HyPhy constants which control optimization options are listed below. The meaning can be looked up in Multimedia File Viewing and Clickable Links are available for Registered Members only!!  You need to Login

OPTIMIZATION_METHOD
OPTIMIZATION_PRECISION
GLOBAL_STARTING_POINT
USE_DISTANCES
MAXIMUM_ITERATIONS_PER_VARIABLE
USE_LAST_RESULTS
RANDOM_STARTING_PERTURBATIONS

HTH,
Sergei

Thanks for the optimisation info. Regarding parameters, I guess I was unclear: what I am doing is to untie topologies over many (20) partitions, so the program has to optimise a large number of branch length parameters. The problem does not appear when I use a single partition, so I suspect it is the simultaneous optimisation of many branch lengths along with the omega distribution parameters that is causing it to go astray.

Konrad

Looks like I was right: using a tighter OPTIMIZATION_PRECISION makes the problem go away in most cases. Using 5 test cases, with the default value of OPTIMIZATION_PRECISION the unconstrained model had lower lnL in 4 of them; tightening the precision by a factor 100 removes the problem for 3 of the 4 (the remaining one ran for ages and still gave a lower lnL).

Konrad

I have followed this thread with great interest and the latest reply
is quite informative. I was also a little worried with some
differences in the results I was obtaining with HYPHY vs PAML.

It would be fantastic if you could post this tests cases you used
somewhere so that this could be given as a proof case scenario.

Thanks in advance,

   Albert.

Dear Konrad and Albert,

One more thing: check MAXIMUM_ITERATIONS_PER_VARIABLE and messages.log for messages to the effect "Optimization routines returned before precision achieved...".

Finally, the functions which converge poorly are usually indicative of flat likelihood surfaces (e.g. overparameterized models), or strange domain artefacts (which usually arise in mixture and constraint problems).

Cheers,
Sergei