AUClimRisk
Bayesian network studio · 0.2.0
Double-click to add · Shift-click to select several. Drag nodes to arrange · Alt-drag to pan · Ctrl-wheel to zoom.
Samples CSV Distribution CSV
Posterior distribution
Summary
All nodes

Observations for the active scenario

Enter observations in a single table, then apply them together. Blank cells remove hard observations. Soft, interval and noisy evidence stay intact unless replaced.

Tornado sensitivity

Uses every state of an unconditioned categorical node, or its 5th, 50th and 95th posterior percentiles. Each setting is recalculated. The dashed line is the active-scenario baseline. Inspect the detailed rows for impossible settings or low sampling precision.

All tested settings and diagnostics
Export sensitivity settings

Strength combines normalised information gain with absolute weighted correlation. It describes posterior association, not a causal intervention effect.

Sensitivity ranking
Metrics

AUClimRisk first infers structural disturbances from factual evidence, then replays those same disturbances after the selected do-interventions.

Factual vs counterfactual
Effect summary

Choose the active scenario in the header.

Choose Add evidence or Set intervention in the Build rail, then click a node. Evidence conditions inference; interventions replace a node's causal mechanism.

Active conditions
Compare scenarios
Export comparison

Choose an action under uncertainty

Mark categorical root nodes as Decisions and numeric expression nodes as Utilities in the inspector. Utilities are added; larger values are preferred. Use negative costs to minimise expenditure. This analyser evaluates a single decision stage.

Expected utility of each action
Value of perfect information
Export decision table

Learn from data and propagate cases

CSV headings match node IDs. Boolean values accept true/false, yes/no or 1/0; categorical values use state values or labels. Blank and NA cells are missing. Extra columns are ignored; .scenario optionally names cases.

Download CSV template
Data preview (first 12 rows)
Learn probability tables

Each node uses rows complete for that node and its parents. Unseen parent combinations retain their existing probabilities. Missing values are not imputed.

Cases and scenarios

Import up to 100 scenarios, or calculate up to 250 cases without adding them to the project. Each case starts with its own observations.

Export case results

Aggregate event frequency and severity

A separate compound-Poisson calculator: independent, identically distributed event losses, independent of annual frequency. All values are illustrative unless calibrated with your data.

Assumptions

Uses the particle count and seed in the Inference panel.

Annual aggregate loss
Risk measures
Export simulated years

A model you can interrogate

AUClimRisk 0.2.0 is a local Bayesian network workbench for exploring uncertainty, diagnosing outcomes and comparing actions.

Start with a five-minute exercise
  1. Load Sprinkler diagnosis from the Case study panel.
  2. Run inference. Wet grass has a prior probability of 64.71%.
  3. Choose Observed wet grass in the scenario selector, then run again. Rain rises to about 70.79%.
  4. Open Scenarios, select several scenarios and Calculate comparison.
  5. Open Sensitivity, choose a target state and run a tornado analysis.
  6. Save Model downloads the whole project, including tables, layout, roles, notes and scenarios.
Build and edit

Add a node or double-click empty map space. Edit its ID, name, type and definition in the inspector. Apply changes commits the edit. In Connect mode, click the parent followed by its child. The model must remain acyclic.

Adding a parent replicates the existing probability rows. Removing a parent averages its rows equally; review the resulting table. Undo reverses committed edits, including layout changes. Ctrl/Cmd+D duplicates selected nodes; Shift-click adds nodes to a selection.

Use Generate table for a uniform table, weighted ranked approximation or Boolean Noisy-OR. Ranked generation uses parent-state midpoints and a truncated Normal on [0,1]; it is an approximation rather than AgenaRisk's exact ranked semantics.

Understand the calculations

Small discrete networks use exact enumeration. Larger finite networks use bounded variable elimination; node marginals are exact, while joint exports use posterior draws. Hybrid networks use likelihood-weighted Monte Carlo. Check effective sample size and warnings before interpreting simulated results.

An observation updates beliefs. An intervention replaces the node's mechanism and cuts incoming causal influence. Observation tornado charts describe conditional sensitivity; intervention tornado charts evaluate causal changes only if the graph and assumptions justify a causal interpretation.

Counterfactuals retain inferred structural disturbances and replay them after an action. This assumes the inverse-CDF structural model used by the app. It is not an identification test for arbitrary causal networks.

Data and decisions

Learning updates existing discrete tables from complete node-family cases with optional Dirichlet prior counts. It does not learn network structure or run EM. Preview the changes before applying. Batch results include errors for impossible cases.

Decision analysis enumerates root decision combinations and maximises the sum of utilities. Perfect information is supported for one categorical uncertainty observed before the decisions. Multi-stage policies and continuous value-of-information optimisation are outside this release.

Save, recover and share

Save Model is the portable, durable copy. The browser also keeps a recovery snapshot for this app address; Restore recovery opens it. Browser storage can be cleared or unavailable, so download important work. Separate tabs keep separate recovery sessions.

Export report creates a printable HTML model specification with available calculated results. Export SVG saves the risk map. JSON files from versions 0.1.0 and 0.1.1 can be opened directly.

Current boundaries

This release does not read AgenaRisk .ast/.cmp/.cmpx files, implement dynamic discretisation, nested risk-object inference, EM learning, or multi-stage influence diagrams. Groups organise the explorer and map; all nodes remain in one network.

AgenaRisk desktop workflow reference

Complete versioned workspace representation used by upload and download.