Dear All,
This message contains replies to a number of recent postings on:
IMPROVINGS BUGS
***************
I'm grateful to Tony O'Hagan and others for suggesting improvements.
Here are a few comments (remember I am not the programmer!):
> 1. What about a list of recently used files in the File menu? This
> is common in Windows software and very helpful.
This and other MS-type facilities would be good, but have not been a
priority up to now. I will check feasibility.
> 2. From BUGS I can save graphics into a compound document or paste
> into something like Word, but how can I edit such graphs? What
> format are they in? Can we have graphs in a format understood by
> other systems, e.g. postscript? Or can we have access to the data
> defining the graph, like coordinates of points on curves?
The graphs, doodles etc can be transferred as OLE objects. This means
that, for example, a Doodle copied to Word can then be opened and
edited just like an Excel spreadsheet - (it is odd to see the WinBUGS
menus working in Word, but this means, for example, that one can change
variable names into longer ones in documentation). Essentially they
can be edited as much as they can in WinBUGS, which for output graphs
is, unfortunately, not at all. One can, however, run WinBUGS from inside
Word, except the size of the object window seems to be restricted at
the moment. Postscript versions of all output etc can be obtained by
printing as if to a postscript printer but really to a file (see
manual).
Values of traces can be obtained from the `coda' button. Values for
autocorrelations and GR plots can be obtained as described in the manual.
Value for desnity plots are not cuirrently available - does anyone want them?
> 3. I can sort of save the whole MCMC output in a form readable by
> CODA, but I don't seem able to recreate it in a later BUGS session.
> So suppose I do a run and save some results like means or density
> plots of some nodes. Then later I want to start up BUGS again and
> look at some other nodes or some other kinds of output like
> quantiles. Because the MCMC run itself is not saved, I can't do
> this. I have to do a fresh run. Not only is this wasteful (and
> perhaps very time consuming), but then the data on which the new
> outputs are derived are different from the data on which the previous
> outputs were calculated. What about an option to save the status
> from one WinBUGS run to another as a kind of workspace?
We decided that saving all simulated values is just not sensible.
Since the same seed for the generator is used each time (unless changed
by the user), then the run can be reproduced exactly. This would, of
course, be helped by a batch run facility (see below). (You can always
use `save state' to save the entire current status and start again
later.)
> 4. A related question. I want to be able to try various different
> versions of a model, perhaps with different priors. I don't seem to
> be able to change any of the data once the model is compiled. I have
> to go back and input the entire model and data (in its modified form)
> from scratch. I understand that recompilation would be necessary,
> but why does BUGS seem to forget everything once it has compiled?
> Can we have the facility to re-open the current model specification,
> input a few changes to data, then recompile? Can we even be
> allowed to reopen the model specification, change the model and then
> not have to input data that have not changed?
I'm afraid I am not qualified to answer this - we'll try to come up with an
answer.
> The old "classic" BUGS had a scheme for putting a sequence of commands into
> a file and running the file. This made it very easy to do new runs, which
> made debugging easy.
This is our most frequent request, and is exactly what I want to do as
well. Our priority now is for a `batch' version of WinBUGS that can
be run on many platforms.
I hope this helps a bit and thanks again for the suggestions. Now I'm off
on holiday!
David Spiegelhalter for the BUGS team
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|