revise the causal explanation notebook, added test #447
Conversation
* added preempt and its tests * refactored test to use new code * lint * added consequent_differs and a test thereof --------- Co-authored-by: eb8680 <eb8680@users.noreply.github.com>
* added preempt and its tests * refactored test to use new code * lint * added Preemptions and a test thereof * added `undo_split` and a test thereof (#422) * import --------- Co-authored-by: Eli <eli@elibingham.com>
* added preempt and its tests * refactored test to use new code * lint * added Preemptions and a test thereof * added `undo_split` and a test thereof * added SplitSubsets and a test thereof * added `undo_split` and a test thereof (#422) * import --------- Co-authored-by: Eli <eli@elibingham.com> Co-authored-by: eb8680 <eb8680@users.noreply.github.com>
rfl-urbaniak
added this to the Causation and explanation components with ac and explanation notebooks milestone
* docstrings update WIP * small typo
* add soft_neq and a few tests in test_ops * sof eq WIP * conversion to soft_eq * defaulting soft_eq scale to .1 * Tavares conditions have landed * soft_neq with tests * added docstring for soft_neq * refactor soft_eq and soft_neq * tests * remove failing tests * comment * sign * move soft_neq to internals for now * nit * remove comment --------- Co-authored-by: Eli <eli@elibingham.com>
* Add explainable module to sphinx build * Add explainable module to sphinx build
* Add explainable module to sphinx build * Add explainable module to sphinx build * reorganize codebase * remove empty file * rename alternatives * reorganize test files * sphinx * remove test_defaults
* allow for empty witness in SearchForCause * removed mwc_empty from a test * allow empty ant and con
There was a problem hiding this comment.
Can you explain the role of the condition_on_consequent argument to ExplanationEvaluation in the notebook, and why it's set to False throughout? I thought you always wanted to condition on the consequents in the factual world, since otherwise there's nothing to explain.
eb8680
added
status:awaiting response
|
Added the following explanation: In principle, we want to be able to condition on the observed consequents: as whatever happened to the consequents is what we want to explain. On the other hand, some of the conditions in the original definition involve queries in the unconditioned model (e.g., are the consequents different always if we intervene on a given candidate cause?) In a more general setting to mimic the definition from the book we would need to have two separate model runs, with and without conditioning and run inference to propagate the results of the conditioning. In this context, when we're dealing with not too many discrete variables, we can proceed in a simplified manner and get away with running forward with an unconditioned model. The conditioning on the consequent is then delegated to trace handling for those parts of the definition that require conditioning. We will emphasize this element of trace handling conditioning in a comment when it happens. |
rfl-urbaniak
added
status:awaiting review
|
@rfl-urbaniak can you resolve these new merge conflicts? |
|
@rfl-urbaniak and @eb8680 , what is the current status of this PR? I have some spare cycles if you'd like my help reviewing. |
SamWitty
added
status:WIP
Updating the notebook in light of the refactoring.