Lecture 2: Abduction and inference to the best explanation

Mars Rover

Back to the Mars Rover example.

These steps do not describe an iterative process. All these steps are intertwined.

1) Generating candidate hypotheses: does not come for free. You cannot just brute force this: a doctor would not check all possible hypothesis; the patient would die before he has the list of candidates.

Pierce's work is also about the possibility of finding new hypotheses in logic.

2) Decide what is relevant in sensory input?

Overload of input; intractable problem.

What feature changes are relevant for a particular goal, in a given environment?

Cf. Frame problem.

3) When to gather new evidence?

4) Select best hypotheses

What is "best"? This is what Lipton's paper is about. Not necessarily the most probably one: e.g. doctor's most probably explanation is "you are ill", but this is not helpful (not specific enough).

5) Learn from experience

6) Try to infer casual relationships and test hypotheses by interventions

Abduction (terminology of this course)

Follows Haselager's book.

In literature different terms are used in various manners, so we agree on the following terminology:

  • Abduction is "the overarching inferential process of explaining (relevant) phenomena we observe"
  • Abduction proper is the sub-process of generating or selecting candidate hypotheses for consideration
    • Not yet choosing one
  • Inference to the best explanation: choosing "the best" of these hypotheses.
  • "Best" remains unclear and to be specified.
  • "Relevant", is info instrumental in the inferential process?

Burks paper:

Recap on terminology:

  • Why are induction and abduction non-demonstrative?

Not logically valid, no logical consequence from the premises: no QED. You cannot demonstrate the logical truth of a statement.

  • Map meaning of abduction in early and later period

Period 1: both induction and abduction evidencing, both former descriptive and latter explanatory. Period 2: abduction generates and induction tests hypotheses.

Insights:

In period 1 both induction and abduction seem forms of "inference to the best explanation", just with a different focus.

Induction in the second period is thus now "inference to the best explanation"! (And abduction specifically to generation only )

  • Do we need a logic of discovery?

Yes, for symbolic AI this would make hypotheses generation programmable. If you assume generating hypotheses is a matter of genius, then your chances of formalizing that and programming it in AI are nihil.

So this old project is actually quite relevant!

Logic is here not necessarily propositional etc., but indicates some form of explicit formalization.

  • Can we formalize this logic of discovery? How do we do this in real life?

E.g. debugging a problem. We do probably not assume the output or some external library is wrong. We would sooner hypothesize our computation is incorrect.

E.g. we have 0.4 x 0.5 = 0.9

Reasoning: what would we expect? What do we have?

H1. we seem to have computed addition. H2: it outputs 0.9 always

How can we formalize coming up with H1 and H2?

Mark Blokpoel: will argue that analogical reasoning underlies this process of hypothesis generation: look at a analogous concept/problem and how that works.

Burks addresses this question when he says that inference is deliberate or conscious controlled. Under this premise you exclude intuition and genius type of inference (because this is not called inference at all!)

The crux is: is intuition part of the reasoning process?

Follow-up: is meta-reasoning about one's own inferential process unique to humans? You could frame this again in another way: matter of AGI.

Harman paper on Inference to the best Explanation

Enumerative induction:

  • Weak: the next swan is white
  • Strong: all swans are white

The induction on swans is not warranted because I know it's unlikely that all animals have the exact same colour. The induction that the sun will rise tomorrow is however warranted.

Claim of this paper:

1) If enumerative induction is warranted : you can also call it "inference to the best explanation" 2) If you cannot, it is probably not warranted and thus not an inference at all.

Problematic sentence: from inference to best explanation, infer truth of hypothesis

Picture of woman having a rose at the ear:

1) Hypothesis 1: she is listening to a rose * Explains picture perfectly! But a priori very unlikely 2) Hypothesis 2: she is part of experiment (my own) * More likely but does not explain this specific picture (why )

So Kwisthout: Harman's phrase does not take a priori likelihood into account!

Def: knowledge as justified true belief

So belief cannot just be true, but also must be justified

  • So not just the truth of conclusion
  • Justification could be authority
  • Also means that the lemmas you use to arrive at knowledge are relevant as part of the justification process
  • The justification is relevant for determining what is the best explanation

Lipton

What is actually best in "inference to the best explanation"

  • Self-evidencing explanation
    • Hypothesis is supported by the evidence it is supposed to explain
    • Because when the hypothesis is true ... FINISH
    • Circularity
    • See paper for "red-shifted" nature of spectrum
    • REVIEW: the circularity

Main point: most probable explanation is not necessarily the best explanation.

The Raven paradox

Two models of epistemology:

1) Deductive-Nomological model 2) Hypothesis-Deductive model

Both are problematic w.r.t. inference to the best explanation.

1) Deductive-Nomological model * Too loose, too permissive. * Not everything you can logically infer from a hypothesis is support for the hypothesis. * Predicted observations that are true are support for hypothesis * But this does not take into account that predictions can be completely irrelevant

H ravens are black P Sun will rise tomorrow

Assume H is true. Then also H or P is true. Observe P. Yay, more support for your hypothesis! But P is completely irrelevant for H!

2) Hypotheco-Deductive model

All ravens are black implies logically, that everything that is not black is not a raven. But seeing something that is not black is not evidence that all ravens are black!

Lipton's own answer: best explanation provides most understanding.

Likely and lovely?

Kwisthout is very sceptical about Lipton's remark that best is "the most understanding likeliest to be correct".

The explanation that would, if correct, provide the most understanding, is the explanation that is judged likeliest to be correct (Lipton 2004)

Counterargument: when you say "patient is ill" this is likely to be correct but not because it offers most understanding, i.e. likeliness and loveliness can strongly diverge.

(REVIEW: I did not fully get this; see lecture notes)

"P, rather than Q" -> what counts as observation?

Your choice for P and its alternative contraints possible explanations.

So the formulation of the observation itself has influence on the possibility of explanation!

REVIEW: why is this important? W.r.t. Lip.

Thermometer example:

1) Observe change in thermometer, e.g. observe: bar rises * then explanation could be: temperature rose 2) Observe that bar stays within limits of the thermometer * This warrants really different explanations than before, while arguably you "see" the same.

Nice example:

Ignaz Semmelweis' 19th century investigation: led him to a contrastive question, i.e. P rather than Q.

(Why do some mothers die and others not?) The relevant contrast: some doctors had autopsy class before.

Glass 2012

Out of time.

Goal: alternative formalization of "best"

Often the most probably explanation is the best one, but not because it's most probably but for another reason.
(More confirmary power?)

Truth tracking.