a simple account of explanations

[tobiasgerstenberg]

here is a simple way to get started:

let's say the only things in our vocabulary are:

• A, B, E
• collided
• and
• not

some rules:

• the faster ball always collided with the slower ball
• we counterfactualize be removing the agent from the scene: if A collided with B, we simulate what would have happened if A had been removed from the scene shortly before that collision
• we only mention events that didn't happen when these would have made a difference to the outcome

some applications:

1. causal chain
   • explanation: "E went through the gate because A collided with B and B collided with E."
   • justification:
   • this is true since if A hadn't collided with B, B would not have moved, and A would not have been there anymore
   • if B hadn't collided with E, then E would not have moved since we remove B, and A is stationary at this time (since it already collided with B)

2. double prevention
   • explanation: "E went through the gate because B collided with A and A did not collide with E."
   • justification: we mention the negative event here (A not colliding with E) since that event would have made a difference to whether the outcome would have occurred

3. preemption
   • explanation: "E went through the gate because A collided with E."
   • justification: we don't mention that B did not collide with E since this event would not have made a difference to the outcome

4. causal chain 2
   • explanation: "E went through the gate because A collided with E."
   • justification: we don't mention that B collided with E since that event made no difference to whether or not the outcome happened

Limitations:

• this simple model cannot deal with situations of overdetermination (like the one below)

• to deal with this case, there are several options:
  1. construct joint events: "Ball E went through the gate because ball A and B together collided with ball E."
  2. include a notion of sufficiency
  3. included how-causation and say that ball A and B affected how ball E wen through the gate.

[joeyginorio]

A few questions:

1. For the causal chain, example 1, the explanation given is: --"E went through the gate because A collided with B and B collided with E."

Do we not want either of these as candidate explanations? --"E went through the gate because A collided with B." --"E went through the gate because B collided with E."

2. If we only generate one explanation per clip, do we then not worry about scoring?

[tobiasgerstenberg]

Here are some qualitatively different kinds of explanations:

• partial explanations (like the ones for the chain you mentioned above)
• overinformative explanations (e.g. "E went through the gate because A collided with B and B collided with E and A didn't collide with E."; it is true that A didn't collide with E but we don't need to mention that since it would not have made a difference to the outcome)
• wrong explanations (stating something that isn't true)

So, good explanations are true and contain only what's necessary for the listener to figure out what happened (no less, and no more). It's these pragmatic principles that we need to consider when scoring explanations.

For informativeness: how much does the explanation help cut down the space of possible things that could have happened. For simplicity: we prefer explanations with fewer cause

[joeyginorio]

So for the qualitatively different explanations, are we going to hand-craft those for each clip? Or should the model have a way to construct/score the whole space of explanations which include the qualitatively different kinds you mentioned?

[tobiasgerstenberg]

It should be fairly easy to generate these:

• we just take all the events that happened, and all the events that didn't happen, and construct all the possible combinations of these events (using AND)
• for example, in the causal chain, one explanation would be: "Ball E went through the gate because ball A collided with ball B and ball A didn't collide with ball E."

[joeyginorio]

Ok, so we want the model to construct/score the space of explanations, including the qualitatively different ones. I think what counts as an explanation is now clear.

I'll make updates to the doc, and follow up when questions come up.

[tobiasgerstenberg]

great! just to clarify: explanations are just statements of the sort "Ball E went through the gate because event 1 and event 2 and ... ." good explanations are those that include the right events.

different models differ in how the choose what events to include. for example, a useful baseline comparison model is one that just always states the events that actually happened. this model:

1. doesn't care about whether the events played a causal role
2. never mentions events that didn't happen

another model might construe events in the wrong way, e.g. when it records a collision that happened between A and B and A was faster than B, it records that "B collided with A". this will lead the model to generate a different counterfactual (it will simulate a counterfactual in which ball B is removed shortly before the collision with A) from a model that encodes the event as "A collided with B".