Open vkhodygo opened 2 years ago
Correction: the original source is inconsistent as well. The text says:
r_i that is drawn from a logistic distribution
Whereas the Algorithm 6
states:
r_i ← a random number from uniform distribution (0,1)
This looks like some sort of a copy-paste error to me. The logistic distribution is bounded meaning 1 / r - 1
is bounded as well resulting in finite values of log(1 / r - 1)
whereas this is not the case for the uniform distribution (0, 1)
. However, the uniform distribution is strictly positive as opposed to the logistic one that is defined on the whole R.
Note, that the original paper makes a clear distinction between open and closed intervals such as (0, 1)
and [0, 1]
. That means using the default rng generating values in the range [0, 1)
is incorrect even in the case of SBDAlignment
.
See original research.
SBDLAlignment
class used to use the same RNG (see also #26) asSBDAlignment
for generation of random numbers before and even now in v5.0.0. This is incorrect,SBDLAlignment
must rely on the conventional logistic distribution.