An old mystery solved

[Julien Yann Dutheil]

Hi There,

For some versions now the optimization reparametrization option did not lead to accurate convergence, even with simple models. And I could not figure out why, until now!

It happened at some time that I added another transform for interval constraint: atanh in addition to tan. the new (hyperbolic) option was favored and made to default. The issue is, however, that there was an error in the implementation of the derivative:

d(tanh(x))/dx = 1 /cosh^2(x),

and this was incorrectly coded as 1/cosh(x^2).

When fixed, we get convergence again! I'm looking forward to play again with the option and will soon publish a new benchmark of bppml's performances :)

Julien.