---

name: inference-sampling-and-simulation description: "Guide pgmpy exact and approximate inference, sampling, simulation, and DBN inference workflows." disable-model-invocation: true metadata: disco-role: operating license: MIT

Inference, Sampling, and Simulation

Use this sub-skill when a task asks for posterior probabilities, MAP assignments, exact or approximate inference, samples from an already-parameterized model, Dynamic Bayesian Network inference, or model-level simulation.

Do not use this sub-skill to build or fit the model first. Route graph/CPD construction and check_model() failures to modeling-and-factors, structure or parameter learning to learning-structure-and-parameters, causal identification, ATE, and effect regressors to causal-identification-and-effects, and model/data loading or metric evaluation to data-io-and-evaluation.

First decisions

1. Confirm the model is parameterized and valid before inference or simulation: call model.check_model() on Bayesian-network-style models after CPDs are attached.
2. Decide whether the user wants an observational query or an intervention:
   • query(..., evidence={...}) and simulate(..., evidence={...}) condition on observed variables.
   • simulate(..., do={...}) generates data under an intervention. If the task asks to identify or estimate a causal effect, route to causal-identification-and-effects instead of treating conditioning as a substitute for do.
3. Use state names exactly as stored in CPDs. If the model was created with named states, evidence such as {"Test": "positive"} is valid but {"Test": 1} may not be.
4. Set show_progress=False or globally disable progress bars with pgmpy.global_vars.config.set_show_progress(False) for automated scripts.
5. Use seed= for repeatable simulation and approximate inference.

Routing table

Need	Primary pgmpy entry point	Notes
Exact posterior on a discrete BN, Markov model, factor graph, or junction tree	pgmpy.inference.VariableElimination or BeliefPropagation	Prefer VariableElimination for one-off exact queries; consider BeliefPropagation when repeatedly querying a calibrated junction-tree workflow.
MAP assignment	infer.map_query(...)	Query variables must not overlap evidence variables.
Large discrete model where exact inference is too expensive	pgmpy.inference.ApproxInference	Sampling-based; increase n_samples and use seed for reproducibility.
Forward, rejection, or likelihood-weighted samples from a BN	pgmpy.sampling.BayesianModelSampling	Direct sampler API; model.simulate(...) wraps it for common workflows.
Gibbs samples from a BN or Markov network	pgmpy.sampling.GibbsSampling	Uses integer state encodings in generated Markov-chain states; provide a valid start state when needed.
Model-level synthetic data with evidence, interventions, soft evidence, or missingness	DiscreteBayesianNetwork.simulate(...) or DynamicBayesianNetwork.simulate(...)	Returns a pandas object or array depending on DBN return_format.
Dynamic Bayesian Network exact inference	pgmpy.inference.DBNInference	Variables and evidence use tuple nodes like ("X", 2).

Bundled references and script

• Inference and sampling API map lists public imports, signatures, return shapes, and automation notes.
• Workflows gives copyable recipes for exact/MAP inference, approximate inference, direct sampling, simulation, soft evidence, missingness, and DBN queries.
• Troubleshooting covers invalid evidence, missing CPDs, query/evidence overlap, exact-inference tractability, progress bars, seeds, virtual evidence, and DBN time-slice conventions.
• inference_smoke.py builds a tiny BN from scratch, validates it, runs exact query/MAP checks, compares belief propagation, and simulates a few rows without reading repository files.

Minimal copyable pattern

from pgmpy.global_vars import config
from pgmpy.inference import VariableElimination

config.set_show_progress(False)
model.check_model()
infer = VariableElimination(model)
posterior = infer.query(
    variables=["Disease"],
    evidence={"Test": "positive"},
    show_progress=False,
)
map_state = infer.map_query(
    variables=["Disease"],
    evidence={"Test": "positive"},
    show_progress=False,
)

If that code fails, inspect the model's CPDs and state names before changing the algorithm. Most inference errors come from an invalid model, evidence variable not in the graph, evidence state not in a CPD, or asking for the same variable in both variables and evidence.