changes

来自「隐马尔可夫工具箱」· 代码 · 共 74 行

Version 2.0 (Current, 24/08/2001):
__________________________________
Version 2.0 features the H2M/cnt  extension for models with Poisson or Negative
binomial state conditional distributions. All  functions in h2m/cnt/ are new as
well as h2m/octave/gammaln.m.

ex_sprec has  been modified so  as to try  to determine automatically  which of
matlab or octave is used.

Also fixed  a long standing problem  that could be annoying  for people reading
the code, although  it didn't change the result:  hmm_fb, hmm_em, hmm_mest used
to contain lines like this one

  % To compute the log-likehood, use: log(sum(alpha(T,:))) + logscale.

However, since we use scaling,  sum(alpha(T,:)) = 1 and thus the log-likelihood
is directly  given by logscale  - which has  been renamed to logl  for improved
clarity.  The term log(sum(alpha(T,:)))  corresponds to the log-likelihood when
scaling is not used  and is a remainder of versions prior  to 1.3 where scaling
was not systematically used.


Version 1.5 (June 8, 1999):
___________________________
Added  ex_sprec and statdis,  renamed doc/examples  to ex_basic  and suppressed
doc/impnois.m.   Added MATLAB  V5  compilation pragmas  in gauseval,  gauslogv,
mix_par and mix_post.


Version 1.4 (December 30, 1997):
________________________________ 
In hmm_tran, the  reestimation of the transition matrix in  the general case is
now much simpler.


Version 1.3   (July  16, 1997):
_______________________________
In hmm_fb, the scaling is now performed at each  step to avoid uneeded troubles
(this only  slightly slows down the  computation). Version 1.1 introduced a bug
(!)  which has been corrected: using the same scale factors (in hmm_fb) for the
alpha and the beta is needed for the transition matrix reestimation formulas to
be correct (see hmm_tran).

An important error was corrected  in gauseval and  gauslogv (the version 1.2 of
these functions yielded incorrect results in the case of scalar observations).

multgen is now  called randindx.  There is a  new function called hmm_psim.  vq
can now be used in the case of scalar observations as well.


Version 1.2  (June  17, 1997):
______________________________
The computation  of gaussian density values  was performed by   a piece of code
that was duplicated in many function. This now  forms a new function gauseval.m
(and gauslogv.m  which  does the same  thing in  log). hmm_fb.m, mix_post.m and
hmm_vit.m have been modified to take this into account. User of MATLAB compiler
should compile gauseval.m  and  gauslogv.m because  it saves  much time.    The
mex-file c_dgaus.c still  exists but is usefull  only for people who don't  use
the compiler (note that the  existence of c_dgaus.mex   is now only checked  in
gauseval.m).


Version 1.1 (May 23, 1997):
___________________________
In hmm_fb, scaling is performed also when the forward variable becomes
too big. In fact when the original data is "small", the forward
variable tends to infinity and not to zero: this could cause overflows
in the original version of hmm_fb.


Version 1.0 (Mar 24, 1997):
___________________________
Corresponds to the original mfiles (no version number).