From: AAAI-87 Proceedings. Copyright ©1987, AAAI (www.aaai.org). All rights reserved.
olecullar colllections
Collins
and Kenneth ID. Forbus
Qualitative
Reasoning
Group
Department
of Computer
Science
University
of Illinois
Abstract
Hayes
has
reasoning
research
identified
about
two distinct
liquids.
has focused
his contained-liquid
a technique
on applying
collection
uids in terms
a system.
ogy, and present
tions
given
“pieces
tained
in MC
that
MC
descripontol-
a Qualitative
stuffs.
how this representation
plex conclusions.
Sometimes
or system
Process
theory
two distinct
For example,
examples
rules
liquid
while
also
Contained-Liquid:
the physical
world.
about
a hypothetical
col-
through the system
Likewise. a river
may be viewed either as a static container of water defined
by its banks (i.e.. the same river it was a century ago), or
collection
each
of
which retains its identity as it flows to the sea.
As Hayes [6, 71 notes, neither ontology alone suffices
to
explain
commonsense
neither
ontology
of little
reasoning
alone
aided engineering.
suffices
pieces
about
of water,
liquids.
Similarly,
for intelligent
computer-
about
“the pressure
at a portal” in the contained-liquid
ontology,
ble to explain the details of a thermodynamic
but impossicycle with-
out following
It is easy to reason
a “piece
piece-of-stzlfiontology,
notions of continuity
of stuff”
This
paper
presents
reasoning with descriptions
We introduce the molecular
590
through
the system.
The
as we shall show, makes explicit the
of space and conservation
of matter,
but provides no mechanism
behavior of the system.
for reasoning
a technique
about
the overall
for generating
and
of fluids as “pieces of stuff”.
collection (MC) ontology as a
Engirwkring Problem Sslvisrg
some open
Hayes ontologies:
the liquid
If the container
Contained
Amount-of
an engineer must think of “the
as an object (the contained-liquid
reasoning
for perform-
and discuss
the original
Consider
single object.
views of an object
about
lection of molecules traveling together
as an object (the piece-of-stuflontology).
as a dynamic
inferences,
in a container
is open then
ble for liquid to leave the container
but interrelated
in the container”
ontology)
rules
and illustrate
their use with several
that this representation
provides a
We begin by reviewing
and discuss
can be used to draw com-
to reason
sometimes
implemented
of con-
its description
model
these
Introduction
are needed
in terms
to computing
problems.
to enter.
I.
We show
We claim
MC descrip-
We illustrate
implemented
terms.
ontology.
on the Contained-Stuff
of a system
is a prerequisite
basis for more complex
liq-
traveling
on the Contained-Stuff
is parasitic
in that a description
stuffs
ing this computation,
examples.
We argue
a specializa-
of stuff”
of Hayes’ piece-of-stuff
the MC ontology
ontology,
using the
which represents
rules for generating
using contained
using several
ontology,
We claim
are parasitic
tions
for
physics
and generalizing
descriptions
(MC)
that
This paper presents
ontology
of little
ontologies
qualitative
ontology.
tion of his alternate
through
Most
for generating
molecular
specialization
which
is emptied
may be influenced
and refilled.
Consider
quantity
by various
pro-
condensation).
They
as when a cup of coffee
In this ontology
of coffee are viewed as the same
Piece-of-Stufl:
and for new liquid
liquids have a continuous
cesses (e.g., flow, evaporation,
may disappear and reappear,
as a
it is possi-
a particular
the two cups
object.
collection
of mole-
cules as a unit traveling around inside a system.
The
collection
of molecules
will have a fixed mass and a
continuous position in space, which is influenced
velocity, which in turn is influenced by various
by its
forces
acting upon the object.
A piece of stuff is never created or destroyed (assuming conservation
of mass), so
there
are fewer problems
of changing
existence
from
this ontology.
It. is straightforward
Liquid
scribes
to
generalize
the
Contained-
ontology into a Contained-Stuff
ontology
that degasses and allows multiple substances
as well .3’.
Qualitative
Process
ate descriptions
theory
of contained
[3, 4, 51 can be used to generstuffs,
and we build on those
descriptions.
In i7j no restriction
is made as to the size of a piece of
stuff.
We obtain the molecular collection
(MC) ontology
by stipulating
that the collection
be so small that we can
assume it is never distributed
over more than one place (we
return to this later). This tiny piece of stuff is viewed as a
collection
of molecules - as opposed to a single molecule
Figure
This schematic
of the importance
about
1: The
SWOS
of a Navy Propulsion
problem
of the MC ontology.
A sophisticated
this system is, “Given an increase
what happens to the steam temperature
Understanding
Figure
plant provides an illustration
in feedwater
temperture,
at the superheater
what happens in this situation
question
outlet?“.
is one of the hardest
2: Sample
developed
with particular
MC movements
processes,
describe
how the
MC'splace and state change as a consequence of that process acting.
Space limitations
preclude showing the entire rule set.
IfFlow (source,
destination,
path)
then if Location(MC.
source>
Transition(MC,
problems given at the Surface Warefare Officers’ School, in New Port,
R.I. The representation
rules for generating
These rules, associated
PLACE,
if Location(MC,
in this paper provides a basis for
Transition(MC,
answering this question.
path)
path)
PLACE,
destination)
IfBoiling(substance, container)
then Transition(MC, STATE, GAS)
IfCondensation(substance. container)
then Transition(MC, STATE, LIqUID)
imply
the direction
in the solution
ontology
-
so that
it may possess
temperature
and pressure.
molecules
such macroscopic
to be considered
Any ontology
For reasoning
as
tion,
collection
of
Contained-Stuff
it is important
the world into individuals:
that
the number
of individu-
als be few. The Contained-Stuff
ontology partitions
system into a few discrete objects using the natural
aries
provided
ontology
by containment.
fails to preserve
individual
molecules
But
molecular
the
Considering
and unnecessary,
of them act more or less alike.
By con-
sidering the possible behaviors of an anonymous collection
of molecules,
we constrain
the possibilities
for the whole
by considering
on the Contained-Stuff
ceeded
in saying
conditions
tology.
anything
for reasoning
We believe
collection
ontology.
coherent
with
ontology
is par-
No one has suc-
about
establishing
the
collection
on-
the molecular
the reason
establishes
for this failure
is that
the
motion.
a global
view
But
of the
method
establishing
physical
for determin-
the gradient
system.
The
provides
this viewpoint
for the MC ontology
by establishing
paths
and conditions
for flows and state
changes.
reasoning
The
with the results
and consequently
sions.
Consider
based
on molecular
of the Contained-Stuff
does not have to re-derive
the system
shown
in Figure
collections
description.
those conclu-
1. Figuring
out
how MC moves requires knowing the mass properties
of the
fluid, viewed with respect to the components
of the system.
Looking
the pressure
solely
at MC, there
differences
between
is no way to establish
system
components
that
must
like flow direc-
be used.
trying
Given
pressure
a
as
to find the pres-
of molecules.
causing
Here, a pump
a liquid
it is in and what
For our purposes,
flow into
is happening
MC is uniquely
the MC history
example
through
generate
the total envisionment
The total
to it
defined by
and interesting,
third
The
critical
the situation
should
how these
is that
of MC's history.
of which
Engine
between
is selected
step finds the possible
observation
The
and the specific
of all consistent
(QPE) to
for which the MC
to be meaningful
locations
each active
(in fluid systems)
some
active
of time.
properties
Processes
situations
them.3
involve
last for an interval
establishes
ifies a fragment
Process
In order for the history
and should
of MC and
in five steps.
for the given configuration.
transitions
a single situation
is desired.
The
consists
by the possible
Second,
occurs
knowledge
the Qualitative
envisionment
connected
requires
Contained-
Stuff ontology
skirts
than
gradient,
of places
Constructing
processes
a local
collection
facts
the place it is in, the substance
of which it is composed,l
and its current phase (i.e., solid, liquid or gas). The place
is the container
or fluid path in which MC resides.2
notion
ing such
rather
collection
a pressure
in those places.
history
provides
play a role
the boiler.
Our goal is to construct a history for MC, describing
MC ontology alone is insufficient.
Global information
is required to identify what MC is doing. In classical physics the
of gradient
MC must
molecular
we can talk about
first step is to feed the domain
only one individual.
We claim that the molecular
asitic
of location
the sequence
Contained-Stuff
identity.
would be prohibitive
since all the billions
a fluid
bound-
ontology
description,
sure on an arbitrary
the
To determine
the Contained-Stuff
a function
as a unit MC.
must divide
Although
problem,
is inadequate.
properties
Call the arbitrary
of flow.
of the
and states
can
change.
process
operate
specon ob-
jects,
some
stuffs.
what,
We can associate rules with each process to describe
if anything,
its activity
implies about the location
‘In this paper only single substance
will be contained
systems are considered.
“Potentially, this could be refined through the use of some coordinate system such as submerged-depth
or the length along a path.
3Each situation r e p resents a unique set of active views and processes taken together with the signs of derivatives for all quantities.
and
phase
liquid-f
source,
of MC. For example,
the
rule
associated
with
low implies that when MC is in liquid form in the
it can move into the path of the flow, and end up
in the destination
Figure
2).
of the flow without
The
rule associated
MC will undergo
same location.
can compute
changing
with
boiling
state
(see
implies
that
a liquid to gas phase transition
within
By combining
these partial histories,
the full spatial
extent
of MC’s travels
associated
phase transitions
(if any) .*
In the fourth step, the Ds values
5
are computed.
By
assumption,
from the surrounding
tained
Changes
in Heat,
Temperature,
For example,
liquid
if the
flow is greater
by rules associated
temperature
than
=
con-
Volume
and
with processes.
of the
destination
the temperature
of a
of the source,
III.
then both Heat and Temperature
of MC will be increasing
in the destination.
During boiling Heat is increasing
and
The MC-history
during
number
condensation
Finally,
the relevant
them.
From
nomena
these
it is decreasing.
the fifth step constructs
places and the possible
this graph
as branching
histories
often
it is easy
the graph defined by
movements
between
to recognize
such
For example,
steam
out of a ship’s boiler is often tapped off for several
purposes, such as driving the propulsion
turbines,
generators
laundry.
to produce
The choice
electricity,
of which
the goal of the reasoning.
path
paths
be considered.
that
MC retains
which
and powering
path
coming
different
running
the ship’s
to take will depend
Sometimes
of a specific
must
phe-
or cycles of flow. In real fluid systems
branch.
on
it is the properties
are of interest.
In other cases all
However, it will be assumed
its identity,
i.e.,
that
the molecules
of MC
never split themselves
between two paths.
This assumption is realistic if one considers MC to be a tiny subset of the
liquid
described
in the Contained-Stuff
view of the same
sys tern.
Two
observations
are relevant
here.
First,
the MC-
history generation
algorithm
is linear in the number of
active process instances,
making it quite fast.”
Second,
the
relationship
and the states
between
the
episodes
of the envisionment
in the MC history
is slightly
complicated.
One state in the envisionment
can give rise to a number of
episodes in the MC history.
For example, the steady flow of
working fluid in a refrigeration
system would typically
described
as a single state in the envisionment
using
Contained-Stuff
ontology.
But
viewed
from
be
the
the MC level,
it will give rise to episodes
involving
heating,
liquid-gas
phase transition,
compression,
gas-liquid
phase transition,
etc.
ness we describe
a.
Ds value of a quantity is the sign of its derivative.
“Running the rules over a total envisionment takes roughly one
minute: constructing an MC history for any situation afterwards takes
5- 10 seconds.
592
Engineering Problem Solving
several
3 illustrates
tainers
choose
connected
from the
liquid
of them
a scenario
in both
Figure
below.
consisting
of two open con-
containers
and the
4 shows the MC history
flow rates
for this situ-
ation. The MC history is annotated
with information
the type of process responsible
for the movement,
as the derivatives
in the history.
complex
on a
For concrete-
by a pump and a return fluid path. We
envisionment
the equilibrium
situation,
exists
have equalized.
has been tested
complexity.
FPow
Figure
where
algorithm
of varying
Pumped
and state
This
examples,
(phase)
information
about
as well
of MC at each
becomes
more
such as in the refrigerator
low. Even though the situation
is in steady
derivatives
in the Contained-Stuff
ontology
place
useful
example
for
be-
state (i.e., ail
are zero), the
MC history shows that each little piece of stuff in the system undergoes continuous
change, both in position and in
pressure.
However, since there is no way for MC to enter or
leave the system, one can conclude that the total amount
of stuff in the system
B.
is constant.
A Refrigerator
One of the motivations
for looking at the MC ontology
was to allow reasoning about complex thermodynamic
cycles such
as that
a simple
refrigerator
used
in a refrigerator.
involving
Figure
the situation
state,
where
denser.
forced
rator.
where
two
and condensation),
As in the pumped
for the MC history
is
through
it returns
MC boils in the evapora-
the compressor
to the liquid
phase
to the conand is finally
through the expansion
valve back into the evapoThis representation
provides the foundation
for an
important
heat
selected
5 shows
processes:
all flows have equalized.
6 shows the MC history.
tor and then is pumped
Figure
six seperate
heat flows, two state changes (boiling
a compressor
flow and a liquid flow.
flow example,
4More than one history can be produced if there are disconnected
components in the fluid system. Each history corresponds to a different choice of subsystem for MC.
generation
of examples
the steady
‘The
example
the
we
(MC)]
inherited
are determined
pumped-flow
for MC’s quantities
Ds [Amount-of
is simply
stuff.
3: A simple
and its
0. Pressure
Height
Figure
during
class of engineering
boiling
and
loses
conclusions.
it during
Since
MC gains
condensation.
it
Figure
4: The MC history
for the pumped-flow
Figure
example
7: The
SWOS
MC history
PWMPEO-FLOW
0
”
0
:,
c
F
L
0
W
k
W
LIQUID-FLOW
Location
Can1
Pump
Can2
F-P
Ds[Heat]
0
0
0
0
0
0
1
0
0
0
0
0
0
0
-1
0
, Ds [Temperature]
I
t Ds [Pressure]
Ds [Volume]
I Ds [Heiahtl
I
u
I
Figure
u
I
u
Location
State
1 Sea 1 Pump
5: A refrigerator
must
Pl
11101
11
Ill11
0 lo
more
heat
via the expansion
I
0
0
0
through
1
1’
1
lo
1 O
Thus
temperature.
this history
ation
the steam
is shorter,
less time
means
the steam
generation
in the superheater,
implying
be calcubased on
temperature
making
The higher steam
heat will be transferred,
plant scenario of
The result of an
can in principle
analysis (DQ)
the increased
episode
spends
the refrigera-
probllem
141. Roughly,
rate higher.
than
is a net heat flow
to the condenser.
The SWOS
’
the compresser
valve, so there
uphill to a higher
that the boiling
MC history
0
0
0
tor is pumping
heat
1 Env
I
-1
-1
-1
increased feedwater temperature
lated by a differential
qualitative
refrigerator
-1
-1
-1
0
0
I
Here we return to the Navy propulsion
Figure 1. Figure 7 shows the MC history.
6: The
1
1
0
1
1
0
I
from the evaporator
6.
Figure
P2
0
0
1
1
I
ILIGIGIGIGIG
S-H
0
0
0
0
0
be moving
returns
B
L
Ds [Heat]
Ds [Templ
Ds[Press]
Ds[VolumelI
Ds[Height]I
IUI
B
IL/
I
gener-
rate means
hence
less
a lower temperature
at
the superheater
outlet.
Weld (in press) describes
a set of
DQ rules which, combined
with this representation,
may
be powerful
enough
The
to reason
ability
provides
to draw this conclusion.
significant
with
ogy. The Contained-Stuff
to determine
termines
tology
Location
State
-__
I Evap
Liq
DsCHeatl
Ds [Temp]
___.
4 Ds[Pressl
1 DsCVol&nel
! DslHeishtl
’
,
’
I
1
-1
0
1
1
Evap
Gas
1
I
i
!
0
0
0
0
1
1 Comp
! Gas
j
I
1
/
1
1
1
1
-1
0
Cond
Gas
!
1 Cond
1 Liq
1 EValve
1
Liq
-1
1
0
-1
(
)
/
1
0
0
0
0
I
-1
!
-1
1
-1
-1
-1
0
0
,
views
of a situation
over using
ontology
a single
provides
which
processes
are active,
the overall
behavior
of the system.
provides
the complementary
ability
ontol-
the conditions
and thereby
The
de-
MC on-
to reason
about
where a piece of stuff came from and where it might go.
We demonstrated
that MC histories can be easily computed
from
’
1
4
4
multiple
advantages
QP
models
of fluids
organized
around
Contained-
Stuffs, and argued that this representation
pro\,ides the
basis for several
important
engineering
inferences
(i. e.,
closed-cycles,
analysis).
recognition
of heat
pumps
and differential
Collins and Forbus
593
It is unclear whether or not growing an MC history
across transitions between situations in the ContainedStuff ontology is a good idea. If one is considering a liquid
system that oscillates, for instance, then this could be necessary. However, most questions that arise in engineering
concerning the MC history are about steady-state behavior,
i.e., a single situation in the Contained-Stuff ontology.
The MC ontology is based on infinitesimal pieces of
fluid. It may be possible to generalize it to spatially extended pieces of stuff. This generalization would provide
the ability to, for example, identify the spread of a contaminate through a fluid system.
We have only begun to explore the reasoning potential of the MC ontology. Currently we are implementing
rules to calculate quantity space information involving MC
parameters. Furthermore, we plan to augment the MC history by associating equations with each movement. These
equations will be combined to yield quantitative descriptions of relevant system parameters, such as efficiency or
work output per pound of working fluid. A differential
qualitative (DQ) analysis could then be performed to identify how these parameters could be optimized, or in general
how a change in one quantity will affect the behavior of the
system.
V.
Acknowledgements
Brian Falkenhainer, John Hogge, and Gordon Skorstad
provided valuable comments, both in developing the theory
and in writing the paper. John Hogge’s ZGRAPH program
has been invaluable for displaying the MC histories and total
envisionment graphs. This research was supported by the
National Aeronautics and Space Administration, Contract
No. NASA NAG 9137, and by an IBM Faculty Development award.
References
[l] de Kleer, J. “An Assumption-Based
nance System”, Artificial Intelligence,
Truth Mainte28, 1986
[2] de Kleer, J. and Brown, J. “A Qualitative Physics
based on Confluences”, Artificial Intelligence,
24,
1984
[3] Forbus, K. “Qualitative Process Theory” Artificial
telligence, 24, 1984
In-
[4] Forbus, K. “Qua 1it a t ive Process Theory” MIT AI Lab
Technical report No. 789, July, 1984.
[5] Forbus, K. “The Problem of Existence”,
of the Cognitive Science Society, 1985.
Proceedings
[6] Hayes, P. “The Naive Physics
systems in the Micro-Electronic
Manifesto” in Expert
Age, D. Michie (Ed.),
Edinburgh University Press, 1979
!7] Hayes, P. “Naive Physics 1: Qntology for Liquids” in
Hobbs, J. and Moore, B. (Eds.), Formal Theories of
the Commonsense
World, Ablex Publishing Corporation, 1985.
594
Engineering Problem Solving
PI Kuipers,
B. “Common Sense Causality:
Deriving
Behavior from Structure” Artificial Intelligence,
24:
1984
Shearer, J., Murphy, A., and Richardson, H. Introduc191
tion to System Dynamics Addison- Wesley Publishing
Company, Reading, Massachusetts, 1967
WI
Simmons, R. “Representing and reasoning about
change in geologic interpretation”, MIT Artificial Intelligence Lab TR-749, December, 1983
PI
Weld, D. “Switching Between Discrete and Continuous Process Models to Predict Genetic Activity”, MIT
Artificial Intelligence Lab TR-793, October, 1984
PI
Williams, B. “Qualitative Analysis of MQS Circuits”,
Artificial
Intelligence,
24,
1984