Closed jdramsey closed 7 years ago
The problem here is that there seems to be too much dependence of the algorithm on the initial W matrix randomly chosen; the results for Ling and Lingam change drastically with every run, from the same data, which is quite unhelpful. It's possible this problem was introduced with changes to the matrix libraries in Tetrad. In any case, there is a new version of FastICA in Matlab, and this can be translated into Java and calibrated against the Matlab version just to make double sure that it's right. It may not be necessary to translate everything from the Matlab version, but as much as possible would be nice.
Kun has a better idea.
See Issue #60 and Issue #61. Need to translate the newest version of FastICA into Java.