Bayesian Optimization outputs and interpretation

A completed job returns a per-parameter posterior summary, the full trace for further analysis, sampler diagnostics, and (for pkpd models) compartmental PK parameters. Use the diagnostics to confirm the fit converged before you trust the parameter estimates.

Output artifacts

The job's top-level metrics also report n_samples, n_chains, all_rhat_ok, min_ess, divergences, and loo_score.

Reading the parameter summary

• mean / sd: the posterior mean and standard deviation for each parameter.
• hdi_3% / hdi_97%: the 94% highest-density interval — the range PyMC's default HDI reports as most credible, not a classical confidence interval.
• ess: effective sample size. Low ESS relative to num_samples * num_chains means the chains are autocorrelated and your effective posterior is thinner than it looks.
• r_hat: the Gelman-Rubin statistic comparing between-chain and within-chain variance. Values close to 1.0 (conventionally <1.01) indicate the chains agree; higher values mean the sampler hasn't converged and the summary shouldn't be trusted yet.

Convergence diagnostics

• divergences: MCMC divergent transitions. Any nonzero count is worth investigating — reparameterize the model, raise target_accept, or increase num_tune before trusting the posterior.
• r_hat_max: the worst R-hat across all parameters; check this before reading individual parameter rows.
• ess_min: the smallest effective sample size across parameters; a low value flags the parameter with the least-reliable posterior.

Model comparison

• loo: leave-one-out cross-validation estimate (when compute_loo is enabled). Use it to compare model fits on the same data — higher is better.
• waic: Watanabe-Akaike information criterion (when compute_waic is enabled), an alternative comparison metric on the same scale question as LOO.

PK/PD-specific outputs

For model_type: "pkpd", the parameter summary includes the compartmental parameters for the chosen pk_model (for example clearance CL and volume of distribution V for a one-compartment model). For a population model (population_model: true), the summary includes both the population-level (fixed-effect) parameters and the per-subject random-effect terms named in random_effects.

Plots

When generate_plots is enabled, plots_zip contains:

• Trace plots: per-chain parameter traces, for a visual convergence check alongside R-hat.
• Posterior plots: the marginal posterior distribution for each parameter.
• PPC plots: posterior predictive checks — simulated data from the fitted model overlaid on the observed data, to sanity-check the model against reality.

Read next