DR-BART Implementation

variance = 'ux':
  SCALE_MIX = TRUE
variance = 'x':
  SCALE_MIX = FALSE

std::vector<vec_d> xinfo;

xi : cutpoints mean
xiprec : cutpoints prec

  Each element in the vector represents a variable.
  Each variable has cutpoints

di : data info
contains the x data and also the y data

pinfo pi: Contains MCMC action probabilities

pbd : prob of birth / death
pd : prob of birth given birth / death

getpb : get probability of birth for a tree
  returns 0 when there is no bottom node to split on
  returns 1 when the tree is empty
  else returns 0.5

Mean trees:

bdhet : birth-death heteroscedastic

• can change a mean tree by spawning / deleting new nodes
• First: Decide whether to birth or to death by using getpb
• if: birth operation; else: death operation
• For the birth operation: 
  • Randomly (uniformly distributed) samples a possible node to split on (from goodbots)
  • Randomy (uniformly distributed) samples a possible variable to split on (from goodvars)
  • Randomly (-||-) samples a cutting point from the variable
  • Then calculates metropolis ratio: alpha
    • Therefore calls: getsuffhet
      • getsuffhet iterates over all samples to get [the numbers of samples for the two leaves (left and right) ???]
        • and the likelihood?
        • Therefore calls bn:
          • Assumption: bn takes long when there are lots of u splits
        • But getsuffhet it seems to ignore the other variables it has not split on?
    • If left or right samples < 5 : then do not birth, because the samples are too few
  • Then randomly samples probability (uniformly distributed): if it is smaller then alpha, then do the birth; else do not do it
• ...

drmuhet : draw mu heteroscedastic (model)

• changes a mean tree t
• sets leave means to means of samples considering the variances(?)
  

Precision trees:

bdprec:

drphi:

• changes a precision tree
• updates leave precisions
• uses the gig_norm function. If I recall correctly, this is an approximation of the normal distribution and was mentioned in the paper