Unit A2.1 Causality
Kenneth D. Forbus
Qualitative Reasoning Group
Northwestern University
Overview
• What is causality?• Design choices for causality in qualitative physics• Using causality
– Example: Self-explanatory simulators
A qualitative physics view of causation
• There are several broadly used notions of causality in reasoning about the physical world
• They can be decomposed by several factors, including– Ontological assumptions: Is there a class of entities that
act as mechanisms in the domain?
– Measurement scenario: What sense of change is being discussed?
Measurement Scenarios affect causality
IncrementalCause precedes effect
Continuous Cause, effect coextensive
Heat flow causesheat of water to rise,which causes temperature of water to rise
Moving soup spoon causesthe napkin to wipe your face
Implications for theories of causal reasoning
• Consider the following:– Causes must precede effects in mechanistic situations,
but causes are temporally coextensive in continuous causation.
– Ontological assumptions used by human experts vary with domain
• cf. use of processes versus components in thermodynamics versus electronics
No single, simple account of causality is
sufficient.
“Gold standard” is psychology, not physics
Causality via Propagation
• Source of causation is a perturbation or input (de Kleer & Brown, 1984)
• Changes propagate through constraint laws
• Useful in domains where number of physical process instances is very large
Mythical Causality
• What a system does between quasistatic states– Extremely short period of time within which
incremental causality operates, even in continuous systems
– Motivation: Capture intuitive explanations of experts about causality in continuous systems, without violating philosophical ideas such as “A Cause must precede its effect”
Implications of causality as propagation
• Identifies order of causality with order of computation.
• No input no causality– Quantitative analog: Simulators like SPICE require an
order of computation to drive them.
Causality in QP theory(Forbus, 1981; 1984)
• Sole Mechanism assumption: 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+
Implications of Sole Mechanism assumption
• All natural changes must be traced back to the action of some physical process– If not so explained, either an agent is involved, or a
closed-world assumption is incorrect• The scenario isn’t fully or accurately known
• The reasoner’s process vocabulary is incomplete or incorrect
• Syntactic enforcement: Direct influences only appear in descriptions of physical processes
• Causal direction in qualitative relations crucial for ensuring correct causal explanations
How directional are causal laws?
• Answer: It depends
• In some domains, clear causal direction across broad variety of situations– cf. engineering
thermodynamics
• In some domains, causal direction varies across broad variety of situations– cf. analog electronics
T =f(heat, mass, …)
V = I * R
Causal Ordering
• Used by H. Simon in economics in 1953
• Inputs– Set of equations
(quantitative or qualitative)
– Subset of parameters identified as exogenous
• Output– Directed graph of causal
relationships
• Method (informal)– Exogenous parameters
comprise starting set of explained parameters
– Find all equations that have exactly one parameter not yet explained.
• Add causal links from explained parameters to the unexplained parameter
• Add unexplained parameter to set of explained parameters
– Continue until exhausted
Tradeoffs in causal ordering algorithm
• Advantages– Can provide causal story
for any set of equations • Assuming well-formed
and enough exogenous parameters
– Causal story can change dynamically if what is exogenous changes
• Drawbacks– Poor choice of exogenous
parameters can lead to psychologically implausible causal stories
• e.g., “the increase in blood sodium goes up, which causes the blood volume to go up.”
– Does not specify the sign of causal effect
Self-Explanatory Simulators
• Idea: Integrate qualitative and numerical representations to achieve– Precision and speed of numerical simulation
– Explanatory power of qualitative physics
• Imagine– SimEarth with explanations
– Interactive, active illustrations in textbooks
– Training simulators with debriefing facilities
– Virtual museum exhibits that you can seriously play with
How self-explanatorysimulators are built
DomainTheory
Scenario
IDE & Tools
Compiled Simulation
SIMGEN Compiler
Runtime
SupportFiles
Students
DomainModeler
Curriculum developer, Teacher, or
student
DomainTheory
Scenario
QualitativeModel
CodeExplanationSystem
QualitativeAnalysis
CodeGenerator
Compiling self-explanatorysimulators
How the explanation system works
• Simulator keeps track of model fragment activity in a concise history– <MFi <start> <end> <T,F>>… At any time tick, can recover full activation structure
• Causal questions answered by– Recovering influence graph from activation structure
– Filtering results appropriately for audience• e.g., thermal conductivity not mentioned in Evaporation
Laboratory
• “Can’t say, don’t tell” policy