CHANGELOG

# 2020-01-29 (3)

A new column of the `calibration` data.frame can be provided to control the
initial value of unknown ages, making it possible to repeat an estimation with 
different starting points in order to spot optimisation issues.

# 2020-01-29 (2)

 I found that some estimations end directly after initialization.
Therefore I replaced the likelihood convergence threshold by an argument
that can be controled (`epsilon`). I also propose to bind the number of
iterations to the returned object.

# 2020-01-29 (1)

I found that some calibration specifications failed, and this is
associated with the filling of `ini.time` at the start: the original
code was initializing unknown dates at the midpoints between
calibrations. However this could fail in the following case:

            /-------C1
       C0--X
            \C2

    <--|-----|-------| age


or `((C1,C2)X)C0);` where X has to be estimated and C0,C1,C2 are calibrated.

if the age of C1 is much smaller than the age of C2, we could end up
with X younger than C2, which causes an error. More precisely, it
happens when (age(C0) - age(C1))/2 > age(C0) - age(C2).

The resulting error is then:

    Setting initial dates...
    Error in ape::chronos(tree2, lambda = 100, calibration = subcalib2) :
      cannot find reasonable starting dates after 1000 tries:
    maybe you need to adjust the calibration dates

As a solution, I changed the iteration order of the paths, so that we
initialise first unknown ages preceeding the closest calibrations (i.e.
oldest). In the example, we start with the path [C0-X-C2] *before* [C0-X-C1].

# 2020-01-29 (0)

Copy `chronos.R` from APE v5.3-7 (<http://ape-package.ird.fr>).