So, I've been screwing around with neural nets, and built a learner with
simple linear regression output nodes. The nature of this beast is you
have to invert a big matrix which is close to singular. Because I am a
big numerics nerd, I remembered to add a small number to the diagonal in
order to get an answer .... buuuut ... QR decomposition is not so good
for this.
I spent a few hours wondering why lapack gels and %./domino were not
giving me a proper inverse, when I remembered that LU decomposition is
the way to go for the "near singular" case.
So, if you ever find yourself adding small diagonals to an array when
'%.' barfs at you, reach for
inv=: [: gesv_jlapack_ ] ; [: =@i.
-it will work better.
-SL
----------------------------------------------------------------------
For information about J forums see http://www.jsoftware.com/forums.htm
