Is there a variant of BART that robustly captures noise that is both heteroscedastic and non-parametric (or has an a-priori unknown parametric form)?
For example, a BART that could fit this test data:
r> plot(1:1000 + rexp(1000) * 100)
... sufficiently well that samples from the model posterior predictive distribution would pass common goodness-of-fit criteria with respect to the sample.