Several approaches to reasoning about actions and change co-exist
at present in the literature. The major divide seems to be between
the situation calculus on one hand, and approaches using explicit
time on the other hand. It may not be easy for the readers of
this literature to see how the different approaches relate, and
what are their respective weaknesses and strengths. Sometimes, it
is even difficult for the researchers in the area to make this
analysis. For example, in this recent KR paper, Ray Reiter writes:
There have been a few earlier papers on formalizing natural actions
and continuous time. Shanahan's approach [30] is embedded
in the event calculus (Kowalski and Sergot [11]);
Sandewall [27] relies on a temporal logic. Accordingly, these
proposals are difficult to compare with ours, based as it is on
the situation calculus.
After a suggestion by Ray, ECSTER invites researches in this area
to an on-line colloquium exchange of views on different
approaches to reasoning about actions and change . The basic idea
is to have a mailgroup which is combined with a lasting on-line
presentation of the accumulated contributions, and with a permanent
publication of the entire debate.
The purpose of the debate is to clarify what are the major
alternative approaches to reasoning about actions and change
in contemporary research, and also to identify and compare
the capabilities and the limitations of those approaches .
Some distinctions will be made already at this point in order
to further define the topic. We propose a distinction on ontological
grounds between situation calculus approaches and
narrative timeline approaches , which are defined as follows:
In narrative timeline approaches, one uses a multisorted logic
where "time" is one of the sorts, and actions are attached to
the timeline using a construct such as Do(s,t,a) , saying
that the action a is performed during the interval starting
at s and ending at t . In situation calculus approaches,
on the other hand, one uses an equally multisorted logic
where "situations", as one of the sorts, form a tree-structured
domain where each situation contains a sequence or other
structure of actions. Thus, to express that the property p
holds when the action a is concluded, a situation-calculus
approach would write something of the form
Holds(p, Result(a,s)) , and a narrative timeline approach
would write something along the lines of
Do(s,t) and Holds(p,t) . In both cases, there are of course
many variants to the theme. One topic for the present debate
is what are the advantages and disadvantages of these two
approaches.
Within each of those approaches, and possibly independently of the
distinction, there are various entailment methods which define
how to obtain the intended conclusions for a given scenario
description. Some of these entailment methods are defined in terms
of preference relations or other selection mechanisms on models,
others are defined in terms of syntactic transformations on the
initially given set of axioms. Chronological minimization of
change is an example of a semantically defined method; explanation
closure is an exampleof a method defined through syntactic
transformations. One topic for the present debate is what are the
presently available entailment methods (including both those that
are defined semantically and those defined syntactically) and
what is known about their properties.
A number of techniques which have been proposed in recent years
have been adopted by several researchers. These techniques
include the use of occlusion , filtering ,
nested circumscription , the release predicate , and
composition of actions , but the list can probably be extended.
One topic for
the present debate is what are these generally used techniques,
and to identify cases where a previously known technique
reappears in new guise or disguise.
The concepts of intended models , and of an underlying semantics
defining the set of intended models, have developed as a way of
characterizing what one expects from a logic of actions and change.
This raises a number of topics for the present debate: what are
appropriate ways of defining intended models; in what sense are
intended models truly "intended"; are there alternative definitions
of intended models and how do they relate to each other.
One noticable phenomenon in recent years has been the appearance
of action description languages , in particular the different
variants of the script-A language. Some questions of debate are:
in what ways are action description languages different from logics
(or are they?); why are there so many action description languages;
and how do action description languages relate to underlying
semantics.
When ramification is addressed, there is an issue between those
methods using minimization of change and those methods that
make use of explicit information about causal directions. What is
true about the capabilities and limitations of these alternatives?
This is already a number of non-trivial questions, but that should
not preclude anyone from also addressing other questions of a similar
character with respect to reasoning about actions and change.