
    o  use better colours

    o  dfbase == 1 means fitting lm with intercept, can we work around?

    o  fire a warning when AIC is called in cases the hat matrix is
       meaningless

    o  check AIC for glmboost with non-gaussian family

    o  check cvrisk for blackboost & bodyfat

    o  update benchmarks in cache with cvrisk

    o  use col = rgb(0, 0, 0, 0.1) for lines in plot.cvrisk

    o  AIC(..., "classical") as default???

    o  use extractAIC instead of AIC

    o  adjust control$M when subsetting
    
    o  names in predict and fitted

    o  multivariate responses

    o  reweight generic?

    o  check loss fuer Poisson!

    o  hatvalues + AIC geht so nur fuer GaussReg!
    o  hat matrices fuer GAMS (Hastie & Tib)
    o  check hatvalues, hatmatrix for boosting with non-gaussian family

    o  weighted median for offset in Huber / Laplace, see Hmisc:::wtd.quantile
