LiMiT · Deterministic workspace

Bayesian updates and diagnostic tests

Update supplied probability models and rank tests by expected information gain.

Provide a complete hypothesis model

The hypotheses must be mutually exclusive and exhaustive with priors summing to one. Likelihoods describe the chance of the evidence under each hypothesis; they are not measured by the tool.

bayes: {"hypotheses":[{"name":"A","prior":0.5},{"name":"B","prior":0.5}],"evidence":[{"name":"observation","likelihoods":{"A":0.8,"B":0.2}}]}

Choose the next informative test

Diagnosis ranks binary tests by expected uncertainty reduction per supplied cost. The highest-ranked test is not necessarily the safest action or the optimal entire investigation. The tool does not run the test.

diagnose: {"hypotheses":[{"name":"A","prior":0.5},{"name":"B","prior":0.5}],"tests":[{"name":"separating test","positiveLikelihoods":{"A":0.9,"B":0.1},"cost":1}]}

Avoid double-counting evidence

Multiple evidence items require an explicit conditional-independence declaration. If that assumption is not appropriate, supply a joint likelihood as a single evidence item instead. Evidence assigned zero probability by every hypothesis cannot produce a posterior.

Read the methods and limitations or try the examples in the workspace.