RE: Perdurantist planning problems (was: RE: RE: SUO: Re: Ballot Comment)
>Dear Pat,
>
>See comments below.
>
>
>Matthew West
>Principal Consultant
>Shell Information Technology International Limited
>Shell Centre, London SE1 7NA, United Kingdom
>
>Tel: +44 20 7934 4490 Other Tel: +44 7796 336538
>Email: matthew.r.west@is.shell.com
>Internet: http://www.shell.com
>
>
> > -----Original Message-----
> > From: pat hayes [mailto:phayes@ai.uwf.edu]
> > Sent: 01 September 2001 00:18
> > To: standard-upper-ontology@ieee.org
> > Cc: West, Matthew R SITI-GREA-UK; apease@ks.teknowledge.com
> > Subject: Perdurantist planning problems (was: RE: RE: SUO: Re: Ballot
> > Comment)
> >
> >
> > Now that Adam and Matthew seem to be converging on a mutual
> > acceptance of a spatiotemporal ontology which descibes change in
> > terms of atemporal assertions about temporal parts, and since there
> > seems to be no chorus of protest from the endurantists at this
> > flagrant disregard for their philosophical scruples (Chris Menzel?
> > Mike Gruninger? Are you there?), allow me to mention one severely
> > practical problem with this approach.
>
>MW: Well I am trying to persuade Adam to include an additional 4D
>view, not to replace the 3D view with a 4D view, though I agree Adam
>is heading in that direction at the moment. Anyway ...
> >
> > There is a well-entrenched tradition of 'action planning' in AI which
> > thinks of a dynamic world as moving through states under the control
> > of 'actions', and does planning by proving that states exist which
> > satisfy certain properties, extracting the sequence of actions to
> > achieve the goal by examining the proof. The oldest form of this is
> > planning in the situation calculus, but basically the same idea has
> > been used in a number of different settings.
>
>MW: Well that sounds familiar - do you mean things like Petri nets? If
>not could you give some references please?
Right, Petri nets are one example. The 'classical' AI example is the
use of 'situations' (= instantaneous world-states, more or less, or
at any rate 'stable' world-states) plus 'actions' (functions from
states to states) to describe possible courses of action. Debugging
program code by simulated evaluation is another.
> > What all these have in
> > common is an assumption that it makes sense to reason about changes
> > in terms of a tree (or directed graph) of states and
> > state-transitions, where the states in the immediate 'future' of a
> > state are the alternative possibilities for the next state. Notice
> > that this picture combines two rather different modal ideas, in that
> > one 'dimension' (following paths in the graph) corresponds to time,
> > while the other (the fanouts from each node) to possibility. This is
> > very difficult to reconcile with the perdurantist (4-d) ontological
> > language, since there is no single spatiotemporal 'envelope' which
> > contains all the various possible futures of a given state.
>
>MW: Oh. I hadn't noticed, and planning is one of the things that is
>important.
>
>MW: Lets get a few preliminaries out of the way. When you say "state"
>and "state-transition" you are not refering to specfic spatio-temporal
>extents, but classes of spatio-temporal extent, and the relations between
>them about possibility are what I would describe as classes of relation
>that determine some rules of what is allowed between members of these
>classes. Right?
Well, the properties of those states that are used to describe them
correspond to classes, right. The reasoners only have access to the
classes, typically, not the actual states.
> >
> > The situation is not impossible, since one can think of the planning
> > process as searching through a space of 'expanding' alternative
> > space-time bubbles, but it gets woefully complex to describe, in
> > painful contrast to the elegant style of function nesting which
> > arises from the older ontological frameworks;
>
>MW: My ignorance is showing here. Could you explain, and perhaps give
>an example of what you mean by "function nesting".
Just the syntactic construction of terms by applying function names, as in
(result a3 (result a2 (result a1 s)))
to denote the state got by starting in state s then doing a1, a2 and
a3 in that order. This is typically how AI reasoners build up answers
to planning challenges. The point being that having a single
expression to denote the final state of an action sequence provides a
syntactic hook to hang the inference process on, and that hook has
been found to be very useful.
>
> > and this inelegance has
> > some drastic and possibly fatal computational consequences, since one
> > cannot rely on unification to automatically keep track of the
> > spatiotemporal-part relationships which arise here in the same way
> > that it automatically creates action-sequence terms; the 'fit' of the
> > tree of possible futures with the tree-structure of action terms is a
> > very valuable byproduct of logical reasoning; for many people, in
> > fact, the main product.
>
>MW: OK now I really need examples.
I meant only that the ease of 'classical' action/planning reasoning
seems to depend on the fact that the terms used to refer to states
themselves have a tree structure (ie a syntax tree) which reflects
the path through the tree of possible futures which the planner is
searching to find a solution. As far as I can see (maybe Im just not
seeing things right) the only way to do this adequately in a general
perdurantist perspective requires one to consider much more
complicated kinds of relationship between spatiotemporal 'parts',
which do not conform to the neat tree-like structure of any
expressions in any formal language. (It would be more like matching
drawings or diagrams than matching tree-like linguistic terms.) Maybe
this is just an inevitable fact of life when doing genuinely
complicated spatiotemporal planning; but if so, it would be useful if
we could retain the simple case for use when we only want to do
simple planning.
> >
> > BTW, this is why I long ago gave up on what was otherwise the very
> > promising 4-d approach to naive physics which I developed in my old
> > 'liquids' paper. If what one wants to do is simply describe a
> > particular timeline, then perdurantism has many advantages. But to
> > reason about a complex system of partially overlapping possibilities,
> > it seems to give rise to many ugly and intractable problems, and the
> > elegance of the endurantist simplicity of the situation calculus and
> > its variations becomes very appealing by contrast.
>
>MW: Well I agree that reasoning about the future is difficult. But I
>hadn't noticed it being easier from an endurantist viewpoint.
Simpler, in the cases where that perspective is a 'natural' way to
think. (Not in general, since in other cases it is much more awkward
to reason in the endurantist way.)
>In fact
>some things became clearer for us when we moved to a perdurantist
>view point, but maybe some different things than you have been looking
>at.
Right, many things are clearer there; but not everything, I think.
Pat
---------------------------------------------------------------------
(650)859 6569 w
(650)494 3973 h (until September)
phayes@ai.uwf.edu
http://www.coginst.uwf.edu/~phayes