How to choose optimizer¶
The default choice of optimization methods in params_file is the following:
- Genetic_algorithm as
Global optimizer, - BFGS_log as
Local optimizer.
We recommend to use BFGS_log as local search algorithm. Genetic_algorithm is also the most efficient method for demographic inference with GADMA.
However, there are special cases when it is better to use Bayesian optimization as a global search algorithm. Usually it is case for demographic inference with more than three populations. More about Bayesian optimization could be found here.
Below there is a table with recomendations about global search algorithm choice. GA stands for the genetic algorithm, BO for Bayesian optimization. Symbol ? means that we have not checked the efficiency for Bayesian optimization in that settings.
| Pop.num.Engine | dadi |
moments |
momi2 |
momentsLD |
|---|---|---|---|---|
| 1 | GA | GA | GA | GA |
| 2 | GA | GA | GA | GA |
| 3 | GA or BO * | GA | GA | ? |
| 4 | BO | BO | GA | ? |
| 5 | BO | BO | GA | ? |
| >5 | ✖ ** | ✖ ** | GA | ✖ ** |
* depends on the time of log-likelihood evaluation. If grid size (Pts) or sample size of data are big then BO is better choice, otherwise GA should be used.
** Not supported