0.2    first version to CRAN
 0.2-1  fixed bugs in docs found by scripts on CRAN (R-devel check)
 0.3    added data/echinacea.txt
              man/echinacea.Rd
              test/tutor.R
        added test in aster.R
        added package vignette
        added examples on all help pages to be used by users
 0.3-1  changed variable "site" to "pop" in echinacea.txt
        corresponding changes in many other files
        \usepackage[utf8]{inputenc} in tutor.Rnw
        many small changes in wording in tutor.Rnw
 0.4    made trust the default optimizer
 0.4-1  removed the timing.pdf vignette, fixed xrefs to trust in Rd files
 0.4-2  added import(trust) to NAMESPACE
 0.5    added 4-th family "two.trunc.poisson"
        actually implemented code for k-truncated Poisson
        put document describing k-truncated Poisson in inst/doc
        new R function rktp that simulates k-truncated Poisson
        cannot implement k-truncated Poisson for arbitrary family
              until families get (non-exponential family) parameters,
              such as k for k-truncated Poisson, r for negative binomial,
              offset for Poisson, n for binomial(n, p), sigma for
              normal-location.  All of that goes on the to-do list,
              along with multivariate families (as per revised version of
              the paper), especially multinomial and normal-location-scale.
        fixed (?!) or at least worked around nasty validation bug in
              predict.aster
 0.5-1  fixed bug in tests/tutor.R caused by bug fix in R.  optim was ignoring
            its maxit argument when method == "CG".  When it stopped ignoring
            it, it failed to converge, because it takes more than 1000
            iterations.
 0.5-2  fixed bug in inst/doc/trunc.Rnw -- must be vignette to pass R CMD check
 0.6    new superfamily system: exponential families with non-exponential
            family hyperparameters
        canonical affine models
 0.6-1  negative binomial distribution
 0.6-2  truncated negative binomial distribution
        fixed bug in 0.6 and 0.6-1 that made them useless with models
            that did not have unrestricted canonical parameter space:
            mlogl could return NaN rather than Inf (which is what it
            is supposed to do) when evaluating at point outside the
            canonical space.  Now hopefully fixed with new code
            in astunco.c and astcond.c.  (TODO: need more tests with
            restricted parameter space.  foobar.R is not enough.  Note
            0.6 and 0.6-1 were not made public, so no one bug CJG saw
            this bug.  Still it does show the tests directory is deficient.)