cognitive modeling - analogical reasoning, problemsolving ... · analogical reasoning,...
TRANSCRIPT
![Page 1: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/1.jpg)
Cognitive ModelingAnalogical Reasoning, Problemsolving, and Learning
Ute Schmid
Kognitive Systeme, Angewandte Informatik, Universitat Bamberg
last change: 8. Dezember 2014
U. Schmid (CogSys) KogMod-Intro 1 / 40
![Page 2: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/2.jpg)
Ubiquitousness of Analogical Thinking
Analogy is a powerful and often used cognitive skill:
Exploiting experience from one domain (”base“)
to explain/predict unknown aspects of or solve problemsin a different domain (
”target“).
Analogy pervades all our thinking, our everyday speech and ourtrivial conclusions as well as artistic ways of expression and thehighest scientific achievements.
Polya, How to Solve It, 1945
U. Schmid (CogSys) KogMod-Intro 2 / 40
![Page 3: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/3.jpg)
Analogy as a Central Research Topic in Cognitive Science
e.g., Gentner, D., Holyoak, K.J., and Kokinov, B. (2001). TheAnalogical Mind – Perspectives from Cognitive Science. MIT Press.
Cognitive Simulation Models (Psychology, AI)
Emotional Analogies, Scientific Discovery (Philosophy, Psychology,AI)
Metaphors and Analogy (Linguistics, Philosophy, Mathematics)
Neurocognitive Basis of Relational Reasoning (Neuropsychology)
Analogy in Education (Psychology)
Analogy in (Software) Engineering (HCI, AI)
...
U. Schmid (CogSys) KogMod-Intro 3 / 40
![Page 4: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/4.jpg)
Example 1: Analogy in Literature
Analogy or proportion is when the second term is to the first asthe fourth to the third. We may then use the fourth for thesecond, or the second for the fourth. Sometimes too we qualifythe metaphor by adding the term to which the proper word isrelative. [...] As old age is to life, so is evening to day. Eveningmay therefore be called, ’the old age of the day,’ and old age,’the evening of life,’ or, in the phrase of Empedocles, ’life’ssetting sun.’
Aristotle, Poetics, chap. 21, “Words”
U. Schmid (CogSys) KogMod-Intro 4 / 40
![Page 5: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/5.jpg)
Example 2: Analogy in History and Politics
End of war in Iraq – End of World War II
Global investor – locust
Die Methode, nach der Analogie zu schließen, ist, wie uberall, so auchin der Geschichte ein machtiges Hulfsmittel; aber sie muß durch einenerheblichen Zweck gerechtfertigt, und mit ebensoviel Vorsicht alsBeurteilung in Ausubung gebracht werden.
Was heisst und zu welchem Ende studiert man Universalgeschichte?Akademische Antrittsrede von Friedrich Schiller
am 26.5.1789 in Jena
Used to transport negative/positive emotions!
U. Schmid (CogSys) KogMod-Intro 5 / 40
![Page 6: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/6.jpg)
Example 2: Analogy in History and Politics
End of war in Iraq – End of World War II
Global investor – locust
Die Methode, nach der Analogie zu schließen, ist, wie uberall, so auchin der Geschichte ein machtiges Hulfsmittel; aber sie muß durch einenerheblichen Zweck gerechtfertigt, und mit ebensoviel Vorsicht alsBeurteilung in Ausubung gebracht werden.
Was heisst und zu welchem Ende studiert man Universalgeschichte?Akademische Antrittsrede von Friedrich Schiller
am 26.5.1789 in Jena
Used to transport negative/positive emotions!
U. Schmid (CogSys) KogMod-Intro 5 / 40
![Page 7: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/7.jpg)
Example 2: Analogy in History and Politics
End of war in Iraq – End of World War II
Global investor – locust
Die Methode, nach der Analogie zu schließen, ist, wie uberall, so auchin der Geschichte ein machtiges Hulfsmittel; aber sie muß durch einenerheblichen Zweck gerechtfertigt, und mit ebensoviel Vorsicht alsBeurteilung in Ausubung gebracht werden.
Was heisst und zu welchem Ende studiert man Universalgeschichte?Akademische Antrittsrede von Friedrich Schiller
am 26.5.1789 in Jena
Used to transport negative/positive emotions!
U. Schmid (CogSys) KogMod-Intro 5 / 40
![Page 8: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/8.jpg)
Example 3: Scientific Analogy
The structure of the hydrogen atomis likethe structure of the solar system.
Rutherford, 1911
Analogy as source of scientific and artistic creativity.
U. Schmid (CogSys) KogMod-Intro 6 / 40
![Page 9: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/9.jpg)
Analogy in Education
Prove: the product of two even numbers is an even number
The product of two odd numbers is an odd number.Proof:
Odd number: even +1 = 2N + 1
Product:(2N + 1) · (2M + 1) = 4NM + 2N + 2M + 1 = 2(2NM + N + M) + 1
Analogy
Even number: 2N
Product: 2N · 2M = 4NM = 2(2NM)
U. Schmid (CogSys) KogMod-Intro 7 / 40
![Page 10: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/10.jpg)
Analogy in Education
Prove: the product of two even numbers is an even number
The product of two odd numbers is an odd number.Proof:
Odd number: even +1 = 2N + 1
Product:(2N + 1) · (2M + 1) = 4NM + 2N + 2M + 1 = 2(2NM + N + M) + 1
Analogy
Even number: 2N
Product: 2N · 2M = 4NM = 2(2NM)
U. Schmid (CogSys) KogMod-Intro 7 / 40
![Page 11: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/11.jpg)
Analogy in Education
Prove: the product of two even numbers is an even number
The product of two odd numbers is an odd number.Proof:
Odd number: even +1 = 2N + 1
Product:(2N + 1) · (2M + 1) = 4NM + 2N + 2M + 1 = 2(2NM + N + M) + 1
Analogy
Even number: 2N
Product: 2N · 2M = 4NM = 2(2NM)
U. Schmid (CogSys) KogMod-Intro 7 / 40
![Page 12: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/12.jpg)
Analogy in Software Engineering
Code reuse vs developing from scratch
Systematic support of reuse can reduce development costs dramatically
Defense Information Systems Agency, 1995
In short, the business case for reuse consists of avoiding 80% of thedevelopment costs for reused components (plus some additionalmaintenance savings) minus the 50% extra it costs to build the reusablecomponent in the first place.
Jeffrey Poulin and Brent Carlson, Computerworld, Feb 2004
U. Schmid (CogSys) KogMod-Intro 8 / 40
![Page 13: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/13.jpg)
Analogy and Similarity
Attributes Relationsmapped to mapped to
target target Example
Mere Appearance Many Few A sunflower is likethe sun
Literal Similarity Many Many The K5 solar system is likeour solar system
Analogy Few Many The atom is likeour solar system
Abstraction Few Many The atom is acentral force system
Metaphor x x She is the sun of my live
(Gentner, 1983, 1997)
U. Schmid (CogSys) KogMod-Intro 9 / 40
![Page 14: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/14.jpg)
Three Kinds of Analogy
Proportional
”A is to B as C is to ?D“
Evening is to Day as Old Age is to Life ↪→ “last part of”
most simple form of analogy: transfer of one relation
Predictive/Explanatory
Carry-over of known principles to a new domain of interest
Rutherford Analogy
in Problem Solving
within domain, use of examples (cf. CBR)
Transfer of a known solution to a new problem
Mathematical/programming problems
U. Schmid (CogSys) KogMod-Intro 10 / 40
![Page 15: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/15.jpg)
Three Kinds of Analogy
Proportional
”A is to B as C is to ?D“
Evening is to Day as Old Age is to Life ↪→ “last part of”
most simple form of analogy: transfer of one relation
Predictive/Explanatory
Carry-over of known principles to a new domain of interest
Rutherford Analogy
in Problem Solving
within domain, use of examples (cf. CBR)
Transfer of a known solution to a new problem
Mathematical/programming problems
U. Schmid (CogSys) KogMod-Intro 10 / 40
![Page 16: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/16.jpg)
Three Kinds of Analogy
Proportional
”A is to B as C is to ?D“
Evening is to Day as Old Age is to Life ↪→ “last part of”
most simple form of analogy: transfer of one relation
Predictive/Explanatory
Carry-over of known principles to a new domain of interest
Rutherford Analogy
in Problem Solving
within domain, use of examples (cf. CBR)
Transfer of a known solution to a new problem
Mathematical/programming problems
U. Schmid (CogSys) KogMod-Intro 10 / 40
![Page 17: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/17.jpg)
Subprocesses in Analogical Reasoning
Overview
Representation
Retrieval
Mapping
Inference/Transfer
Learning
(in all cognitive theories: Gentner, Holyoak, Keane, ...)
U. Schmid (CogSys) KogMod-Intro 11 / 40
![Page 18: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/18.jpg)
Representation
Structural representation (graph, term, semantic net, ...)
Problem: Representation crucial for mapping success
on(a,b) vs. below(b,a); x > y vs. x − y > 0
Planet−j
O
O
S
S
O
O
SS
OSR E
CAUSE ATTRACTS
REVOLVES ATTRACTS ATTRACTS MORE MASSIVE HOTTER
O
O
S
S
O
O
S
S
REVOLVES ATTRACTS ATTRACTS
Base
Planet−i
SunYELLOW
Target
Electron−i
Nucleus
MORE MASSIVE
U. Schmid (CogSys) KogMod-Intro 12 / 40
![Page 19: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/19.jpg)
Retrieval
Bottleneck of analogy
Governed by superficial similarity (feature based measures)
Novices fail to identify useful base problem(experiments by Novick, 1988)
In Education: Present suitable base problems explicitely
In Computation: Use structural similarity for retrieval
U. Schmid (CogSys) KogMod-Intro 13 / 40
![Page 20: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/20.jpg)
Mapping
Core of analogy
Structure preserving mapping
First order (same relations/functions in both domains) or higher order
In well-structured, formal domains: homomorphism
U. Schmid (CogSys) KogMod-Intro 14 / 40
![Page 21: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/21.jpg)
Homomorphism
Structure preserving mapping f : S → T
such that f (oS(s1, . . . , sn)) = oT (f (s1), . . . , f (sn))
T TT
oSS
ff =
o
S n
n
S1
T1
x ...
x ...
U. Schmid (CogSys) KogMod-Intro 15 / 40
![Page 22: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/22.jpg)
Homomorphism
Structure preserving mapping f : S → T
such that f (oS(s1, . . . , sn)) = oT (f (s1), . . . , f (sn))
2T T
f =
o
S 2oS S
f
T
S1 x ...
T1 x ...
U. Schmid (CogSys) KogMod-Intro 15 / 40
![Page 23: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/23.jpg)
Homomorphism
Structure preserving mapping f : S → T
such that f (oS(s1, . . . , sn)) = oT (f (s1), . . . , f (sn))
S
f=
oS2
2
S
f
T ToT
S1 x ...
T1 x ...
U. Schmid (CogSys) KogMod-Intro 15 / 40
![Page 24: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/24.jpg)
Example 1: Proportional Analogy
ff =
Day Evening
Old Age
last part of
Life
U. Schmid (CogSys) KogMod-Intro 16 / 40
![Page 25: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/25.jpg)
Example 2: Proportional Analogy
Relation between string concatenation and length of strings
1 x 2N N
x1 2W W
ff =
N
W
+
°
U. Schmid (CogSys) KogMod-Intro 17 / 40
![Page 26: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/26.jpg)
Example 2: Proportional Analogy
Relation between string concatenation and length of strings
x3 2
ff =
°
5
themethe x me
+
U. Schmid (CogSys) KogMod-Intro 17 / 40
![Page 27: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/27.jpg)
Transfer
Based on mapping
Carry-over of information from base to target
”Inference“ of unknown characteristics of target
Transfer/adaptation of a solution
U. Schmid (CogSys) KogMod-Intro 18 / 40
![Page 28: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/28.jpg)
Transfer
Based on mapping
Carry-over of information from base to target
”Inference“ of unknown characteristics of target
Transfer/adaptation of a solution
U. Schmid (CogSys) KogMod-Intro 18 / 40
![Page 29: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/29.jpg)
Example – Transfer
Planet−j
O
O
S
S
O
O
SS
OSR E
CAUSE ATTRACTS
REVOLVES ATTRACTS ATTRACTS MORE MASSIVE HOTTER
O
O
S
S
O
O
S
S
REVOLVES ATTRACTS ATTRACTS
Base
Planet−i
SunYELLOW
Target
Electron−i
Nucleus
MORE MASSIVE
U. Schmid (CogSys) KogMod-Intro 19 / 40
![Page 30: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/30.jpg)
Learning
Acquisition of more general schemes or rules by abstraction (Solarsystem, atom ↪→ central force system)
Analogy as “beginner’s strategy”: Acquistion of general conceptsmakes analogical reasoning obsolete!
U. Schmid (CogSys) KogMod-Intro 20 / 40
![Page 31: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/31.jpg)
Computational Cognitive Science Systems
Forbus, Falkenhainer & Gentner (1989): SMENaive physics
Hummel & Holyoak (1997): LISAProblem solving (between domain)
Anderson & Thompson (1989): modified ACTProgramming/geometrical proofs (within domain)
U. Schmid (CogSys) KogMod-Intro 21 / 40
![Page 32: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/32.jpg)
Cognitive AI Systems
Evans (1968): Geometrical analogies (intelligence test)
Veloso & Carbonell (1993): Plan construction (Prodigy)
O’Hara (1992): geometrical analogies (PAN)
Hofstadter (1995): letter strings (Copycat)
A B C
1 2 3 4 5
U. Schmid (CogSys) KogMod-Intro 22 / 40
![Page 33: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/33.jpg)
Cognitive AI Systems
Evans (1968): Geometrical analogies (intelligence test)
Veloso & Carbonell (1993): Plan construction (Prodigy)
O’Hara (1992): geometrical analogies (PAN)
Hofstadter (1995): letter strings (Copycat)
::::
: :: :
U. Schmid (CogSys) KogMod-Intro 22 / 40
![Page 34: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/34.jpg)
Cognitive AI Systems
Evans (1968): Geometrical analogies (intelligence test)
Veloso & Carbonell (1993): Plan construction (Prodigy)
O’Hara (1992): geometrical analogies (PAN)
Hofstadter (1995): letter strings (Copycat)
GLUE
ROT[p,60] ROT[p,60]
ROT[p,60]
GLUE
ROT[p,60] ROT[p,60]
ROT[p,60]
GLUE
U. Schmid (CogSys) KogMod-Intro 22 / 40
![Page 35: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/35.jpg)
Cognitive AI Systems
Evans (1968): Geometrical analogies (intelligence test)
Veloso & Carbonell (1993): Plan construction (Prodigy)
O’Hara (1992): geometrical analogies (PAN)
Hofstadter (1995): letter strings (Copycat)
abc : abd :: kji : ??
U. Schmid (CogSys) KogMod-Intro 22 / 40
![Page 36: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/36.jpg)
Cognitive Plausibility of Approaches
Unflexible representation(additional re-representation mechanisms in PAN and Copycat)
Mapping of graphs = subgraph isomorphism problem (NP-hard)(heuristics in SME and LISA)
Transfer only”carry-over and replace“,
no real adaptation (permutation, deletion)
Learning not addressedor modelled by an additional mechanism (ACT)
U. Schmid (CogSys) KogMod-Intro 23 / 40
![Page 37: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/37.jpg)
Our Approach
Anti-Unification (Reynolds, 1970): Mapping via common abstraction
Mapping is governed directly by the common structure of base andtarget
Allows use of equational theories for re-representation (Schmid, Gust,Kuhnberger, Burghardt, 2003)
Defined for first- and second-order case (Hasker, 1995)
U. Schmid (CogSys) KogMod-Intro 24 / 40
![Page 38: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/38.jpg)
AU Example
t1 = attracts(sun, planet-i)
t2 = attracts (nucleus, electron-i)
can be generalized tot = attracts(central-body, orbiter)where ‘central-body’ and ‘orbiter’ are variables
Calculating the abstraction results in the mappingϕ = {(sun,nucleus) 7→ central-body,(planet-i,electron-i) 7→ orbiter}Thereby, the necessary substitutions of variables byconstants are also known:σ1 = { central-body 7→ sun, orbiter 7→ planet-i } andσ2 = { central-body 7→ nucleus, orbiter 7→ electron-i }Now the mapping of sun onto nucleus can beperformed by applying first ϕ and then σ2.
t2t1
σ1
σ1’’σ
σ
σ2
2
’
t’
t
U. Schmid (CogSys) KogMod-Intro 25 / 40
![Page 39: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/39.jpg)
First-Order AU Modulo Equational-Theories
Syntactic AU of 10 + (15 · 10) and (20.2 · 3.5) + 3.5 is x + y
The fact, that each term is the addition of a constant and a productcontaining this constant got lost in generalization!
Equational theories can be used to model knowledge aboutre-representations in a natural way
Simple Theory
x + 0 =E xx + y =E y + xx +(y +z) =E (x +y)+z
x + (y * x)
3.5 + (20.2 * 3.5)10 + (15 * 10)= (15 * 10) + 10 = (20.2 * 3.5) + 3.5
{x−>3.5, y−>20.2}{x−>10, y−>15}
U. Schmid (CogSys) KogMod-Intro 26 / 40
![Page 40: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/40.jpg)
Solving Proportional Analogies with E-Generalization
A : B :: C : D
[A]E [C ]E
[A]τ1E [C ]τ2E[A]τ1E ∩ [C ]τ2E
GAC :=
[B]E
[B]τ1E=: Q
[D]E
τ1 τ2
τ−11 τ−1
2
τ−11
τ2
(BA Stephan Weller)
U. Schmid (CogSys) KogMod-Intro 27 / 40
![Page 41: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/41.jpg)
2nd Order Anti-Unification
1st order AU of 10 + (15 · 10) and 3.5− (20.2 · 3.5) is x
The fact, that each term is built by an operation over a constant anda product got lost in generalization!
2nd order AU can be used to model generalization overfunction/predicate symbols
Result: x F (y · x)where x and y are object variables and F is a function variable
Applied to programming by analogy (Wagner and Schmid) in contextof our work on inductive program synthesis techniques
U. Schmid (CogSys) KogMod-Intro 28 / 40
![Page 42: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/42.jpg)
Programming by Analogy
Fac-Problem: If the factorial of 3 is calculated as 3 · 2 · 1 · 1what is the factorial for a natural number n?
Fac-Solution: fac(n) = if(n=0,1,n · fac(n-1)).
NSum-Problem: If the neg. sum of 3 is calculated as ((0− 1)− 2)− 3 what isthe neg. sum for a natural number n?
U. Schmid (CogSys) KogMod-Intro 29 / 40
![Page 43: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/43.jpg)
Programming by Analogy
Fac-Problem: If the factorial of 3 is calculated as 3 · 2 · 1 · 1what is the factorial for a natural number n?
Fac-Solution: fac(n) = if(n=0,1,n · fac(n-1)).
NSum-Problem: If the neg. sum of 3 is calculated as ((0− 1)− 2)− 3 what isthe neg. sum for a natural number n?
Fac-Unfolding:if n = 0 then 1else if n = 1 then 1 · 1else if n = 2 then 2 · (1 · 1)else if n = 3 then 3 · (2 · (1 · 1))
NSum-Problem:if n = 0 then 0else if n = 1 then 0− 1else if n = 2 then (0− 1)− 2else if n = 3 then ((0− 1)− 2)− 3
U. Schmid (CogSys) KogMod-Intro 29 / 40
![Page 44: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/44.jpg)
Programming by Analogy
Fac-Unfolding:if n = 0 then 1else if n = 1 then 1 · 1else if n = 2 then 2 · (1 · 1)else if n = 3 then 3 · (2 · (1 · 1))
NSum-Problem:if n = 0 then 0else if n = 1 then 0− 1else if n = 2 then (0− 1)− 2else if n = 3 then ((0− 1)− 2)− 3
Fac-NSum-Generalization:if n=0 then xelse if n=1 then 1 F xelse if n=2 then 2 F (1 F x)else if n=3 then 3 F (2 F (1 F x))
σfac = {x 7→ 1,F 7→ (· π1, π2)}σnsum = {x 7→ 0,F 7→ (− π2, π1)}Note that arguments of subtractionop are reversed!
U. Schmid (CogSys) KogMod-Intro 29 / 40
![Page 45: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/45.jpg)
Programming by Analogy
Fac-NSum-Generalization:if n=0 then xelse if n=1 then 1 F xelse if n=2 then 2 F (1 F x)else if n=3 then 3 F (2 F (1 F x))
σfac = {x 7→ 1,F 7→ (· π1, π2)}σnsum = {x 7→ 0,F 7→ (− π2, π1)}Note that arguments of subtractionop are reversed!
The abstract term captures the role of 1 and 0 as neutral element and of · and −as ‘combination-operator’ respectively.
Obtaining the target solution: by applying the found substitutions to the recursivesolution of the base problem
Fac-Solution: fac(n) = if(n=0,1,n · fac(n-1)).NSum-Solution: nsum(n) = if(n=0, 0, nsum(n-1) − n).
U. Schmid (CogSys) KogMod-Intro 29 / 40
![Page 46: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/46.jpg)
Evaluation
AU is applicable to model all kinds of analogy
By including equational theories, background knowledge can beincluded in a natural way
Mapping via abstraction vs. direct mappingI Computational advantage
(getting rid of subgraph isomorphism problem)I Cognitive plausibility
(mapping guided by the common role of objects)
Re-representation is modeled by taking into account all possiblerepresentations wrt a given theory simultaneously (investigation ofre-representation triggers in human reasoning by Schmid and Jira)
Learning by abstraction is a side-effect of analogical reasoning
U. Schmid (CogSys) KogMod-Intro 30 / 40
![Page 47: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/47.jpg)
Empirical Evidence
As much as needed vs. as much as possible
Analogy by abstraction and SME: transfer of largest commonsubstructure (“systematic”) (vs. LISA, Copycat and others,“pragmatic”)
Analogy by abstraction: eager generalization during analogy making
Assessing time of generalization (during or after mapping andtransfer) is problematic
Experiment by Eva Wiese (2007)
Domain: Physics, water flow and current flow
U. Schmid (CogSys) KogMod-Intro 31 / 40
![Page 48: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/48.jpg)
Water Flow and Current Flow
U. Schmid (CogSys) KogMod-Intro 32 / 40
![Page 49: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/49.jpg)
Abstraction of Ohm’s Law?
Tutorial: Water Flow (base analogy) or Current Flow (no analogy)
Type-1 Problems: Amperage (I) and Voltage (V)
Type-2 Problems: Amperage (I) and Resistance (R)
Type-3 Problems: Interrelation of I, V, R; Ohm’s Law
U. Schmid (CogSys) KogMod-Intro 33 / 40
![Page 50: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/50.jpg)
Design of the Study
E1 E2 K1 K2
Tutorial Water Currentand Control Questions
Test 1 Type-1 Type-2 Type-1 Type-2
Test 2 Type-2 Type-1 Type-2 Type-1
Test 3 Type-3 Type-3 Type-3 Type-3
U. Schmid (CogSys) KogMod-Intro 34 / 40
![Page 51: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/51.jpg)
Hypothesis: Transfer as much as possible
Time
x
O
O Analogy
Control
Test 1 Test 2
x
Analogical problem solving for I and V facilitates problem solving for I andR (likewise: I and R to I and V)
U. Schmid (CogSys) KogMod-Intro 35 / 40
![Page 52: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/52.jpg)
Hypothesis: Analogy triggers Generalisation
Time
Analogy Control
Test 3
Analogical problem solving triggers generalization over structural relations(I, V, R)
U. Schmid (CogSys) KogMod-Intro 36 / 40
![Page 53: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/53.jpg)
Result Transfer
Mean time differences between Test 1 and Test 2 (p < 0.001); averagesolution time per problem type about 20 sec
Type 1 (I, A) Type 2 (I, V)
Experimental -1,191 -0.979
Control +1,385 +1,557
No speed-accurcy trade-off
Average number of correct answers 88%
Experimental group: 2% more wrong answers in Test 2
Control group: 1% more wrong answers in Test 2
No correlation between solution time and correct answers
U. Schmid (CogSys) KogMod-Intro 37 / 40
![Page 54: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/54.jpg)
Result Generalization
No systematic difference between Experimental and Control Group
Experimental group about 1 sec slower than control group, about10% more wrong answers
Kind of problems used for Type 3 may be unsuitable
New experiment: try to come up with a third domain
U. Schmid (CogSys) KogMod-Intro 38 / 40
![Page 55: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/55.jpg)
Application Potential
Analogy in Education
Systematic support of acqusition of general principles/concepts byproviding useful (structural similar but not identical) examples(Schmid, Wirth, Polkehn, 2003, Cognitive Science Quarterly)
Support for understanding/discovering physical principles (Gust,Kuhnberger; Wiese)
U. Schmid (CogSys) KogMod-Intro 39 / 40
![Page 56: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/56.jpg)
Application Potential
Analogy as weak method for AI systems
Planning: search for a plan has exponential effort in the general caseAlternative: Adaptation of existing plans
Program synthesis: adaptation of existing programs
U. Schmid (CogSys) KogMod-Intro 39 / 40
![Page 57: Cognitive Modeling - Analogical Reasoning, Problemsolving ... · Analogical Reasoning, Problemsolving, and Learning Ute Schmid Kognitive Systeme, Angewandte Informatik, Universit](https://reader031.vdocuments.mx/reader031/viewer/2022022021/5b9fcde309d3f2da5b8be1c7/html5/thumbnails/57.jpg)
Final Remarks
Analogy as an example for interdisciplinary researchin higher cognition
We investigate constraints and processesof human analogical reasoning
We design “cognitive plausible” and formally sound algorithms
Possible applications in many areasfrom education to AI technology
U. Schmid (CogSys) KogMod-Intro 40 / 40