[Eeglablist] runica() result reproducibility improved?

[Delorme, Arnaud]

Hi Makoto,

Your solution of increasing the number of steps and dividing by five the default learning rate (at the cost of having the algorithm running about 5 times slower, right?) is an interesting approach. We will need to check if this is reproducible across datasets with a different number of channels, and can possibly change the default learning rate and stopping rule for the runica algorithm. Also, I assume that by iteration (the rows in your plotted figure), you mean the number of times you run the algorithm to completion (because you also mention 1300 iterations to converge, but these are different types of iterations).

The ordering of the components is based on the amount of variance they explain about the data, so the solution of decreasing the initial learning rate seems to produce more stable solutions in terms of explained variance. 

Note that another way to get reproducible results is to reset the random number generator before running ICA because runica will randomize blocks (type "rng(1);”). However, I agree that this is not an elegant solution and that getting more stable results is essential.

Arno