hi prio, little wrk:
- C: organize .h/.c -files like you're supposed to 
- quantile based knots for P-splines
- discard chains that got stuck (?)
- use rlecuyer for parallel RNG
- plotting for fct:num interaction fails if original data frame has fct's that are saved as numerics...
 since geom_rug uses the original data and then the check in ScaleDiscrete fails :: 
 convert data frame used for plotting purposes accordingly
- use proper dependence on irlba now that it's on CRAN   

hi prio, much wrk:
- vignette:  add more examples on popular datasets?
- clean up weights/scales confusion for gaussian to allow for weighted regression 
- plot mrf with igraph/Rgraphviz? can we do it in ggplot2?
- implement semiparametric terms that are "always included" (i.e. setting w=1 for some terms)  
- intercept in every update block / update intercept with each block 

lo prio, little wrk:
- restartChains
- implement missing value handling
- cyclic splines

lo prio, much wrk:	 
- hierarchical constraints on gamma: 
	sample gamma with constraints (ising prior?)?  
	allow gammas of main effects to influence interactions as well?  (what if main effect has many interactions - dimensionality problem i.e. same gamma applies for >5 alphas?) 
- predict for mrfs in regions w/o observations in original data 
- center higher order interactions not only against main effects but also against lower order interactions?
- don't update iwls weights in every iteration? 
- improve plots for lin (just plot estimated beta, not eta) and its interactions 

eventually:
- redo ssGAMdesign completely formula/parser based (NO fucking grep)
- use Rcpp for updaters
- sparse matrix updates for rnd/mrf ?
- do updates cycling over terms instead of over parameter blocks? 
	(i.e. update (gamma_1, tau_1, beta_1)-->...-->(gamma_p, tau_p, beta_p) instead of (gamma_1,..,p)-->(tau_1,..,p)-->(beta_1,..,p) )  
	if we do this, use compressed designs with unique values only as in bayesX/mboost?
	will probably mean more autocorr since we take smaller steps on the posterior
- improve mixing by switching parameter blocks between chains? 
	(i.e. every 500 iterations, exchange parts of parameter vector between chains...)


