In most computerized aids that quantify uncertainty,
inference is equated with an essentially linear process, in
which a model or knowledge base is built, numerical inputs assessed
(e.g., weights on rules, prior probabilities of facts), and
outputs generated. This linear approach omits the thinking processes
by means of which an analyst selects one consistent set of probabilistic
beliefs (including weights on rules and prior probabilities)
out of all those that are possible.
The Non-Monotonic Probabilist
The Non-Monotonic Probabilitist defines reasoning more realistically.
Actual reasoning about uncertainty in real-world tasks is
typically highly iterative. The results of one line of reasoning
are compared with the results of other lines of reasoning
(or with direct judgment). If there is a discrepancy, the
inputs, parameters, and even the structure of the model or
knowledge base may be revised. The Non-Monotonic Probabilist
provides direct support for the intelligent construction and
modification of inference models in the light of experience
with their application. Reasoning is no longer the blind application
of a fixed model, but a creative cycle in which such models
are built and improved with experience.