Updater Box

Purpose

The Updater box is where you apply updating procedures to an existing model in a Tetrad project. It takes an Instantiated Model (IM) or an Estimator output as input and lets you perform probabilistic learning by specifying:

Values for variables (evidence/observations), and
Manipulations (interventions / do-operations).

Given an instantiated model and those settings, the Updater computes updated (posterior) quantities for variables of interest.

Currently there are two main families of updaters:

Discrete updaters for discrete Bayesian models:
- Junction Tree Updater (default)
- Approximate Updater
- Row Summing Updater
A SEM Updater for linear Gaussian SEMs.

Typical workflow

Connect an instantiated model or estimator
- Place an Instantiated Model or Estimator box on the workbench.
- Configure it to produce either:
  - A discrete Bayes model, or
  - A linear SEM.
- Place an Updater box and draw an arrow from the IM/Estimator box to the Updater box.
Open the Updater box
- Double-click the Updater box to open its interface.
- The interface shows:
  - A list of variables from the input model,
  - Fields to set values (evidence) for some variables,
  - Controls to mark variables as manipulated (intervened on) and assign intervention values,
  - A control for choosing the updater type (for discrete models).
Choose an updater type
- For discrete models, select one of:
  - Junction Tree Updater – exact inference using a junction tree (default).
  - Approximate Updater – sampling- or approximation-based inference for larger models.
  - Row Summing Updater – table-based inference by summing over CPT rows.
- For linear SEM models, the SEM Updater is used.
Specify evidence and manipulations
- For each variable you want to condition on:
  - Enter an observed value (evidence).
- For each variable you want to intervene on:
  - Mark it as manipulated and specify the intervention value (e.g., do(X = x)).
Run the update
- Use the updater’s controls (e.g., an Update or Compute button, depending on your version) to perform inference.
- The Updater computes:
  - Posterior distributions for discrete variables, or
  - Conditional means and variances for SEM variables, given the specified evidence and manipulations.
Inspect results
- Updated probabilities or conditional summaries are shown in the Updater interface.
- You can adjust values and manipulations and recompute to explore “what-if” scenarios.

Updater types and detail pages

The Updater box supports the following updater types, depending on the model type:

Model type	Updater option	Detail page
Discrete Bayes model	Junction Tree Updater	`Tetrad Interface → Junction Tree Updater`
Discrete Bayes model	Approximate Updater	`Tetrad Interface → Approximate Updater`
Discrete Bayes model	Row Summing Updater	`Tetrad Interface → Row Summing Updater`
Linear SEM	SEM Updater	`Tetrad Interface → SEM Updater`

See the individual detail pages for conceptual descriptions and implementation notes for each updater.

Connecting the Updater with other boxes

The Updater box fits into the broader workflow as follows:

Inputs
- Instantiated Model or Estimator:
  - Provides the model to be updated (discrete Bayes model or linear SEM).
- (Sometimes) Data:
  - Data are typically used upstream to estimate the model; the Updater itself works with the instantiated model.
Outputs
- The Updater does not change the model structure or parameters; instead, it:
  - Computes posterior distributions or conditional summaries,
  - Displays them in its own interface.
- You can use these results to:
  - Interpret effects of interventions,
  - Compare outcomes across different evidence/manipulation scenarios,
  - Inform further modeling decisions.

You can also create multiple Updater boxes attached to the same instantiated model to explore different evidence and intervention scenarios in parallel.

Common patterns & tips

Use a discrete updater (Junction Tree, Approximate, or Row Summing) when your instantiated model is a discrete Bayes network.
Use the Junction Tree Updater when exact inference is feasible; switch to the Approximate Updater when models become too large or dense.
Use the Row Summing Updater when you want a conceptually simple, table-based calculation (useful in teaching or debugging).
Use the SEM Updater for linear Gaussian SEMs to compute:
- Conditional means and variances under evidence,
- The effect of interventions (\text{do}(X = x)) in the linear SEM setting.
When performing “what-if” analyses:
- Keep the instantiated model fixed,
- Duplicate an Updater configuration and vary only the evidence or manipulations,
- Compare the resulting updated quantities.