QUAIL '97 (Question of the Day) |
|
Hi. Eventually, after over a week and a half, the results for the QOTD for
the Saturday a week and a half ago. Only Patrick answered this one, so
I put both mine and his here.
I found Patrick's answer to be much more enlightning than mine. Anyway, before
answering the question I had no clue as to the answer, and now at least I
do have one :-)
>
> What ways of Model formation are there in Qualitative Reasoning?
>
Eyal
====
Model formation in QR means that we try to form some "rules" and histories
for the examined system. This is modeling in the "model based reasoning" sense.
We try to form the exact axioms that describe the system. Note that we start
with a given domain theory, and reduce it to the specific task at hand.
P.12: "Models are created from domain theories, which describe the kinds of
entities and phenomena that can occur in a physical domain. A domain theory
consists of a set of model fragments, each describing a particular aspect of
the domain. Creating a model is accomplished by instrantiating an appropriate
subset of the model fragments."
1. Instantiate every law in the domain theory with the possible objects.
This way we can avoid quantities mentioned, and in a simple consise
domain theory, this is practical and simple.
2. Search the space of modeling assumptions (formulas). For each such chosen
set of assumptions we use the previous method. One way to realize
this method is using an ATMS to create all the possible modeling
assumptions combinations.
There are other additional sub-methods that add constraints over the domain
theory, so that the process of model formation will be more tractable.
Patrick
=======
[Little more than a paraphrase of Forbus!]
The compositional modeling approach of Falkenheimer and Forbus holds that
a domain theory (which describes entities and interactions) is made up of
model fragments, and model formation is the process of choosing which
fragments from the domain theory to instantiate.
The simplest approach is to instantiate all fragments. This has the
disadvantage of being very resource-intensive, and is impractical for
large worlds, and not possible when the system is intended simultaneously
to represent alternate perspectives of the same fragment.
Better systems search modelling space to find a necessary set of
fragments to instantiate. Falkenheimer and Forbus use an algorithm that
tried to find all legal combinations of fragments that would produce a
model that could answer a query. Unfortunately this becomes expensive as
the size of the model increases, and generally there may be an
exponential number of such collections of fragments for a problem.
Nayak has a system whereby the domain theory is divided into 'assumption
classes' each of which has an ordering on models into causal
approximations. This produces locally simple but not necessarily
globally simple models; however, it does operate in polynomial time.
Model formulation is usually done iteratively, moving from more crude
generalizations to more specific formulations as the need for precision
or detail increases; currently many models are assembled 'by hand' though
the hope is that soon it will become a fully automated process.
Back to the Question of the Day
Page