Bayesian Abductive Reasoning

A good explanation needs more than a story that fits the facts: it also needs a plausible starting probability. Suppose hoofbeats could come from a horse or a zebra; if both predict the sound equally well and horses are much more common locally, the horse is the more probable explanation. Bayesian reasoning combines these starting beliefs with how expected the evidence would be under each explanation, while abductive reasoning helps us consider possible explanations. Reflect on this: what evidence would make you change your mind about the explanation you currently favor?

Consider explanations before choosing one

Abductive reasoning concerns finding and assessing explanations. Bayesian reasoning updates probabilities as evidence arrives. They are related tools, not interchangeable labels or a guarantee that every appealing explanation is probable. A practical approach is to list plausible candidates, ask what each predicts, and compare those predictions with the evidence.

Weigh two different questions

A prior is how plausible a hypothesis is before the evidence under consideration. A likelihood describes how expected that evidence would be if the hypothesis were true. The posterior is the updated probability after combining both. An explanation can fit the observations well and still be less probable than another if its starting probability is much lower. Conversely, sufficiently diagnostic evidence can overcome an initially low, nonzero prior.

Make the distinction concrete

Imagine a device fails to start. A flat battery and a faulty switch could both explain that observation. The relative frequency of these faults helps set starting beliefs, while an additional test that the switch hypothesis predicts poorly can shift the balance toward the battery. This hypothetical example illustrates belief updating; the best diagnosis depends on the actual tests and alternatives, not simply on choosing the most common fault.

Look for evidence that separates candidates

Instead of collecting details that fit your favorite explanation, ask which observation would be more expected under one candidate than another. In a hypothetical email filter, the frequency of spam provides starting information and patterns in the message provide additional evidence. Treating a single word as conclusive would ignore uncertainty and the possibility of legitimate messages using it.

Keep the hypothesis set open

The probabilities compare the candidates in the model. A missing explanation or a badly specified likelihood can make the conclusion misleading. Simplicity can guide a search for explanations, but a simpler story is not automatically true. Revise both the candidate set and your assumptions when evidence points to a problem.

This lesson combines hypothesis comparison with probabilistic updating, without treating all explanatory reasoning as Bayesian calculation.

• Tenenbaum et al. (2011) describe Bayesian models in which a posterior depends on the product of a prior and a likelihood, normalized over the considered hypotheses; their review explains how structured prior knowledge can support learning from sparse data. [1]
• Douven's Abduction entry (2011 archive) distinguishes explanatory inference from deduction and discusses different ways of formulating inference to the best explanation. Its treatment emphasizes that the precise form and normative standing of abduction remain contested. [2]

Priors and likelihoods can be misspecified, and the available hypotheses may omit the true cause. Incorrect priors do not inevitably reinforce error: sufficiently informative evidence can substantially change a nonzero prior. Under ordinary conditioning within a fixed model, a hypothesis assigned probability zero remains at zero when conditioning is defined, so revisiting the model can matter. A high posterior is relative to the assumptions and candidates being compared, not a proof that the explanation is true. Exact calculations can become computationally demanding; a rough intuitive comparison should not be presented as a calibrated numerical posterior.