BART in the page

y ~ Normal(mu, sigma), mu = BART(x), sigma ~ HalfNormal(1), data sin(x) + N(0, 0.2), sampled by PyMC's own CompoundStep under Pyodide: bartrs's particle Gibbs on the trees, NUTS on sigma, every log density compiled with mode="WASM". The first run installs PyMC, which takes tens of seconds.

Built on pymc-bart and bartrs, with bartrs built for Pyodide from a fork that adds two changes not yet upstream: PGBART compiling its logp in the mode compile_kwargs gives, and a getter for the current forest. The data is synthetic.

points trees draws tune draws not loaded

▬ mu at the current draw; after sampling, its posterior mean 90% interval of mu over the draws so far 90% posterior predictive, mu plus noise - - sin(x) ● data

tree 0

what it adds to mu

The trees: mu is their sum, and each step regrows a tenth of them. Pick one to follow it.

changed shape since the last frame ● leaf ● split