 Notes mostly to myself	{Martin Mchler}
----------------------------------------

1. `bug' (not really : just needs large options(expression = 20000)
    prt.vvec  {but that's an experimental function}

  look at
     example(summary.vlmc)
     print(svb, vvec.p = TRUE)

2. *do* write
     as.dendrogram.vlmc()  and  then  plot.vlmc() from it

   -- partly done: want "hang = 1" argument (we now do +-  "hang = -1").


6. do AIC cutoff search internally -- or at least returning all
   cutoffs (+ df + loglik ?) where pruning happens.

7. Eliminate the maximum alphabet size of 26
   --> src/vlmc.h : max_alpha_len (now 27)  :
      but even better have no fixed size this will make the   struct _node
      much smaller for e.g.the binary case!
   also '27%s' in src/io.c  -- here and other places rely on alphabet of
   only 1 character  (-> allows 255 different ones).
   Better: arbitrary alphabet; keep the alphabet 'vector' (in R sense, can
   have more than one letter) in one place.

8. Allow to update an already fitted vlmc
   by additionally fitting a new discrete time series.
   If you can do this, apply this to allow input of a list of time series
   (for fitting **one** VLMC).

   This should be pretty straight forward, apart from the first few cases
   that have no full context

9. Multivariate VLMC --- new modeling

10. Simulation for model selection, using [AB]IC, plus "Crossvalidation" etc
    --> provide user functions (? or use library(boot) ?

11. Allow to restrict the maximal 'depth' !!


-- vlmc(.. dump =TRUE) and (.. debug = TRUE) should probably NOT use
   std.error, but rather stdout.


----------


predict.vlmc( method = "prob") :
	      First row, give the marginal (estimated, i.e. empirical)
	      probabilities, instead of NAs.

	      --> hence also for "class" etc (instead of NA)

=== predict.vlmc() should allow longer output than input for
    "prediction into the future". --- pretty easy

---------

A. Fit one VLMC using several/many "parallel" series

   definite need; not too hard to program!
   {cheap trick: concatenate these and pad with NA!? -- fails too}

   ==> See more in ./multi-vlmc.R
		   ~~~~~~~~~~~~~~


B. Fit a "multivariate" VLMC:

   Martin's 1st idea: = univariate series in product space.
   Drawbacks:
        - Past is the same length for all series
	  (x1_t may need more looking back of x1 than of x2)
	- Product space is quickly too large  (K ^ p), and quite a few of
	  these states may not appear in data --> not in fitted values.
	  If we'd predict univariately, new multi-states could appear
