Optimize() error: "Initial value and slope must be finite"

[stst]

That worked - thanks for your help! ConjugateGradient is indeed significantly faster than GradientDescent (but still much slower than the NNLS solver I posted in this topic).