Upload
daniel-jack-peters
View
216
Download
0
Tags:
Embed Size (px)
Citation preview
Reminder • Midterm is two weeks from today (Oct
21st!)– Closed book, closed notes, no computers or
calculators or electronic devices– Will cover all material till then.
Overview
• Why qualitative reasoning?• Principles of qualitative representation and
reasoning• Using qualitative reasoning in cognitive modeling
What is qualitative physics?
• Formalizing the intuitive knowledge of the physical world– From person on the street to expert scientists and
engineers
• Developing reasoning methods that use such knowledge for interesting tasks.
• Developing computational models of human commonsense reasoning
Example
• What can happen when you leave a tea kettle on a stove unattended for an hour?
T(s) T(w)
This conclusion can be reached without numerical parameters or differential
equations!
Knowing what might happen suggests when
more precise knowledge is needed
Example
• Suppose we have two identical ice cube trays.One tray is filled with cold tap water.The other tray is filled with warm tap water.We place both trays in a freezer at the same time.Q: Which will freeze first, the warmer water or the cooler water? And why?
Why do qualitative physics?
• Understanding the mind– What do people know? Physical, social, and mental
worlds.
– Universal, but with broad ranges of expertise• Unlike vision, which is automatic
• Unlike medical diagnosis
Why do qualitative physics?
• Can build useful software and systems– Intelligent tutoring systems and learning environments
– Engineering Problem Solving
• Diagnosis/Troubleshooting
• Monitoring
• Design
• Failure Modes and Effects Analysis (FMEA)
– Robots
– Models for understanding analogies and metaphors
• “Ricki blew up at Lucy”
Human-like understanding of complex systems requires qualitative models
Monitoring, diagnosis, failure modes and effects analysis,
creating control software,explanation generation,
tutoring…
Example: Understanding Engineering Analysis
• Important in understanding engineering, scientific reasoning
• Useful model: Solving textbook problems
Cycles
Uniform-StateUniform Flow
Initial State Final State
Single State Missing Information
First Thermodynamics Course
AirT= 25CP= ?
TurbineW=?Air
Air flow
Heater
Turbine
Cooler
Qualitative reasoning in engineering analysis: Example
• No mass flows in or out of the system
• Volume can increase until stop is reached
• Spring force will add to air pressure when it contacts the top
• The temperature will rise, or stay the same if saturated
Heat
Key Ideas of Qualitative Physics
• Quantize the continuous for symbolic reasoning– Example: Represent numbers via signs or ordinal
relationships– Example: Divide space up into meaningful regions
• Represent partial knowledge about the world– Example: Is the melting temperature of aluminum
higher than the temperature of an electric stove?– Example: “We’re on Rt 66” versus “We’re at Exit 42
on Rt 66”
• Reason with partial knowledge about the world– Example: Pulling the kettle off before all the water
boils away will prevent it from melting.– Example: “We just passed Exit 42, and before that was
41. We should see 43 soon.”
Comparing qualitative and traditional mathematics
• Traditional math provides detailed answers– Often more detailed than
needed– Imposes unrealistic input
requirements
• Qualitative math provides natural level of detail– Allows for partial
knowledge– Expresses intuition of
causality
F = MA
A Q+ F
A Q- M
Traditional quantitativeversion
Qualitative version
Example 1: How far will it roll?
• Infants sees baseline sequence of events – medium cylinder rolls
towards wheeled toy, collides, toy rolls some distance
• Infant sees new sequence of events– Like before, but with larger
or smaller cylinder
– Possible or impossible outcome
– Infants look longer event violates their expectations
Example 2: Understanding Science books
• Example: First field guide to weather, National Audubon Society
• “Water vapor is an invisible gas in the air. The amount of water vapor, which is called the relative humidity, can be anywhere between 0 and 100 percent. A relative humidity of 100 percent means that the air is saturated – as full of water vapor as it can be. Most water vapor comes from evaporation. When the sun heats the liquid water in the earth’s oceans, lakes, and rivers, some of it changes into water vapor and rises into the air. The atmosphere also gets small amounts of water vapor from plants and from moisture in soil.”
Qualitative Process Theory
• Key ideas (Forbus, 1981, 1984):– Quantity space: Ordinal relationships between
parameters provide a useful qualitative representation of numerical values
– Influences: A causal, qualitative mathematics that supports partial information provides a natural way to encode relationships involving continuous parameters
– Physical processes: Physical processes provide a notion of mechanism for natural changes.
Quantity Space
• Value defined in terms of ordinal relationships with other quantities
• Contents dynamically inferred based on distinctions imposed by rest of model
• Can be a partial order• Limit points are values
where processes change activation
Twater
Tboil
Tfreeze
Tstove
Examples of limit points in text
• “The temperature at which condensation occurs is called the dew point.”
• “The amount of water vapor, which is called the relative humidity, can be anywhere between 0 and 100 percent. A relative humidity of 100 percent means that the air is saturated – as full of water vapor as it can be.”
Qualitative proportionalities
• Examples– (qprop (T ?o) (heat ?o))– (qprop- (acceleration ?o) (mass ?o))
• Semantics of (qprop A B) f s.t. A = f(…,B,…) f is increasing monotonic
in B– For qprop-, decreasing monotonic– B is a causal antecedent of A
• Implications– Weakest causal connection that can propagate sign
information– Partial information about dependency requires closed
world assumption for reasoning
Qualitative proportionalities in text
• “The more air there is, the more it weighs and the greater its pressure”– (qprop (weight ?air-mass) (n-molecules ?air-mass))
– (qprop (pressure ?air-mass) (n-molecules ?air-mass))
• “As the air temperature goes up, the relative humidity goes down.”– (qprop- (relative-humidity ?air-mass) (temperature ?air-mass))
• Source: Weather: An Explore Your World ™ Handbook. Discovery Press
Correspondences
• Example:– (correspondence ((force spring) 0) ((position spring) 0)
– (qprop- (force spring) (position spring))
• Pins down a point in the implicit function for the qualitative proportionalities constraining a quantity.
• Enables propagation of ordinal information across qualitative proportionalities.
Correspondences as explanation• (correspondence
((distance-rolled bug) Initial) ((size cylinder) Medium))(qprop (distance-rolled bug) (size cylinder)))
• (< (size cylinder) Medium) predicts(< (distance-rolled bug) Initial)
Initial
Model Fragments
• Encode conditions under which domain knowledge is relevant– Participants are the individuals and relationships that
must hold before it makes sense to think about it
– Conditions must be true for it to hold (i.e., be active)
– Consequences are the direct implications of it being active.
• (defmodelFragment saturated :participants ((am :type air-mass)) :conditions ((= (relative-humidity am) 100-percent) :consequences ((saturated am)))
Physical Processes
• A kind of model fragment• But also has direct influences, which are
constraints on derivatives• Examples:
– “Most water [in the air] comes from evaporation. When the sun heats the liquid water in the earth’s oceans, lakes, and rivers, some of it changes into water vapor and rises into the air”
– (I+ (water-vapor am) (rate evap))(I- (amount-of water-body) (rate evap))
– N.B. accumulating bodies of water into an abstract entity, based on shared properties. This is a transfer pattern of influences.
Semantics of direct influences
• I+(A,b) D[A]=…+b+…• I-(A,b) D[A]=…-b+… • Direct influences combine via addition
– Information about relative rates can disambiguate
– Abstract nature of qprop no loss of generality in expressing qualitative ODE’s
• Direct influences only occur in physical processes (sole mechanism assumption)
• Closed-world assumption needed to determine change
Example of a physical process(defModelFragment heat-flow :subclass-of (physical-process) :participants ((the-src :type thermal-physob) (the-dst :type thermal-physob) (the-path :type heat-path :constraints ((heat-connection the-path the-src the-dst)))) :conditions ((heat-aligned the-path) (> (temperature the-src) (temperature the-dst))) :quantities ((heat-flow-rate :type heat-flow-rate)) :consequences ((Q= heat-flow-rate (- (temperature the-src) (temperature the-dst))) (I- (heat the-src) heat-flow-rate) (I+ (heat the-dst) heat-flow-rate)))
Causality in QP theory(Forbus, 1981; 1984)
• All causal changes stem from physical processes
• Changes propagate from quantities directly influenced by processes through causal laws to indirectly influenced quantities
• Naturally models human reasoning in many domains (i.e., fluids, heat, motion…)
F G
Liquid FlowF G
Amount-of(Wf)
Level(Wf)
Pressure(F)
Q+
Q+
I-
Amount-of(Wg)
Level(Wg)
Pressure(G)
Q+
Q+
I+
Time and change
• Time individuated by changes in qualitative state
• Qualitative states differentiated by– Set of active physical
processes– What dynamic
relationships hold– Quantity space values
Spring state
Block velocity
Event structure in QR and NL
• Hypothesis: QR decomposition of event structure is tightly coupled with event structure in NL semantics
• Similarities– Both carve continuous change into symbolic,
meaningful chunks
• Differences– QR suggests finer-grained decomposition on physical
grounds
– NL also decomposes based on non-physical aspects
– QR captures more possibilities
Stopping,
dynamic friction
Stopping, staticfriction
Qualitative states and transitions
Many dynamical properties of systemscan be reasoned about based on
graph properties of envisionments
Loop in states oscillatio
n
Spring-block
oscillator envisionme
nt
Useful properties of first-principles qualitative simulation
• Handles incomplete/inexact data– Easier to extract qualitative data via perception
• Supports simple inferences• Explicit representation of causal theories
– To prevent melting, remove kettle from stove
• Representation of ambiguity– We easily imagine multiple alternatives in daily
reasoning
• Caveat: There are some complexities when viewed as a psychological model
Problems with standard QR as psychological model
• Exclusive use of 1st-principles domain theory – inconsistent with psychological evidence of
strong role for experience-based reasoning
• Exponential behavior of 1st-principles qualitative simulation – inconsistent with rapidity & flexibility of
human reasoning
• Generates more complex predictions than people report – logically possible, but physically
implausible
???
But, QR representations are psychologically plausible(once assumption of exclusive 1st-principle reasoning
is removed)
• Evidence: Protocol analyses– Kuipers, Bredeweg & deKonig, Forbus & Gentner, ...
• Examples– Signs, quantity space representations of number
– Symbolic descriptions of shape and space
– Influences, causal relationships
– Organizing abstractions: e.g., devices, processes
– Relevance conditions, modeling assumptions (for experts at least)
RememberedSituation retrieved from
long-term memory
Qualitative simulation via analogical reasoningObserved
behavior
Analogical inference
provides prediction
Memory matched with
previous situation
Ruling out 1st principles predictions via analogical
reasoning
T(w) = T(fire)
T(w) constant
T(w) increasingT(w) = TBoil(w)
Never seen it happen