l3 apprentissagesebag/slides/cachan_1_2013.pdf · exam final: i questions i problems volunteers i...
TRANSCRIPT
![Page 1: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/1.jpg)
L3 Apprentissage
Michele Sebag − Benjamin MonmegeLRI − LISV
23 janvier 2013
1
![Page 2: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/2.jpg)
Where we are
Ast. series Pierre de Rosette
Maths.
World
Data / Principles
Naturalphenomenons
Modelling
Human−relatedphenomenons
You are here
CommonSense
2
![Page 3: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/3.jpg)
Where we are
Sc. data
Maths.
World
Data / Principles
Naturalphenomenons
Modelling
Human−relatedphenomenons
You are here
CommonSense
3
![Page 4: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/4.jpg)
Harnessing Big Data
Watson (IBM) defeats human champions at the quiz game Jeopardy (Feb. 11)
i 1 2 3 4 5 6 7 81000i kilo mega giga tera peta exa zetta yotta bytes
I Google: 24 petabytes/day
I Facebook: 10 terabytes/day; Twitter: 7 terabytes/day
I Large Hadron Collider: 40 terabytes/seconds
4
![Page 5: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/5.jpg)
Types of Machine Learning problems
WORLD − DATA − USER
Observations
UnderstandCode
UnsupervisedLEARNING
+ Target
PredictClassification/Regression
SupervisedLEARNING
+ Rewards
DecideAction Policy/Strategy
ReinforcementLEARNING
5
![Page 6: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/6.jpg)
Supervised Machine Learning
World → instance xi →Oracle↓yi
MNIST Yann Le Cun, since end 80s
6
![Page 7: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/7.jpg)
The 2005-2012 Visual Object Challenges
A. Zisserman, C. Williams, M. Everingham, L. v.d. Gool
7
![Page 8: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/8.jpg)
Supervised learning, notations
Input: set of (x, y)
I An instance x e.g. set of pixels, x ∈ IRD
I A label y in {1,−1} or {1, . . . ,K} or IR
Pattern recognition
I Classification Does the image contain the targetconcept ?
h : { Images} 7→ {1,−1}
I Detection Does the pixel belong to the img of targetconcept?
h : { Pixels in an image} 7→ {1,−1}
I SegmentationFind contours of all instances of target concept in image
8
![Page 9: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/9.jpg)
Supervised learning, notations
Input: set of (x, y)
I An instance x e.g. set of pixels, x ∈ IRD
I A label y in {1,−1} or {1, . . . ,K} or IR
Pattern recognition
I Classification Does the image contain the targetconcept ?
h : { Images} 7→ {1,−1}
I Detection Does the pixel belong to the img of targetconcept?
h : { Pixels in an image} 7→ {1,−1}
I SegmentationFind contours of all instances of target concept in image
8
![Page 10: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/10.jpg)
Unsupervised learningClustering http://www.ofai.at/ elias.pampalk/music/
9
![Page 11: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/11.jpg)
Unsupervised learning, issues
Hard or soft ?
I Hard: find a partition of the data
I Soft: estimate the distribution of the data as amixture of components.
Parametric vs non Parametric ?
I Parametric: number K of clusters is known
I Non-Parametric: find K(wrapping a parametric clustering algorithm)
10
![Page 12: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/12.jpg)
Unsupervised learning, 2
Collaborative Filtering
Netflix Challenge 2007-2008
11
![Page 13: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/13.jpg)
Collaborative filtering, notations
Input
I A set of users nu, ca 500,000
I A set of movies nm, ca 18,000
I A nm × nu matrix: person, movie, ratingVery sparse matrix: less than 1% filled...
Output
I Filling the matrix !
Criterion
I (relative) mean square error
I ranking error
12
![Page 14: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/14.jpg)
Collaborative filtering, notations
Input
I A set of users nu, ca 500,000
I A set of movies nm, ca 18,000
I A nm × nu matrix: person, movie, ratingVery sparse matrix: less than 1% filled...
Output
I Filling the matrix !
Criterion
I (relative) mean square error
I ranking error
12
![Page 15: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/15.jpg)
Reinforcement learning
13
![Page 16: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/16.jpg)
Reinforcement learning, notations
Notations
I State space SI Action space AI Transition model p(s, a, s ′) 7→ [0, 1]
I Reward r(s)
Goal
I Find policy π : S 7→ A
Maximize E [π] = Expected cumulative reward
(detail later)
14
![Page 17: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/17.jpg)
Some pointers
I My slides:http://tao.lri.fr/tiki-index.php?page=Courses
I Andrew Ng courses:http://ai.stanford.edu/∼ang/courses.html
I PASCAL videoshttp://videolectures.net/pascal/
I Tutorials NIPS Neuro Information Processing Systemshttp://nips.cc/Conferences/2006/Media/
I About ML/DMhttp://hunch.net/
15
![Page 18: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/18.jpg)
This course
WHO
I Michele Sebag, machine learning LRI
I Benjamin Monmege, LISV
WHAT
1. Introduction
2. Supervised Machine Learning
3. Unsupervised Machine Learning
4. Reinforcement Learning
WHERE: http://tao.lri.fr/tiki-index.php?page=Courses
16
![Page 19: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/19.jpg)
Exam
Final:
I Questions
I Problems
Volunteers
I Some pointers are in the slidesMore ?here a paper or url
I Volunteers: read material, write one page, send it([email protected]), oral presentation 5mn.
17
![Page 20: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/20.jpg)
Overview
Les racines : IA
IA as search
IA and games
Promesses tenues ?
What’s new
18
![Page 21: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/21.jpg)
Roots of AI
Bletchley
I Enigma cypher 1918-1945
I Some flaws/regularities
I Alan Turing (1912-1954)and Gordon Welchman: theBombe
I Colossus
19
![Page 22: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/22.jpg)
Dartmouth: when AI was coined
We propose a study of artificial intelligence [..]. The study is toproceed on the basis of the conjecture that every aspect oflearning or any other feature of intelligence can in principle beso precisely described that a machine can be made to simulate it.
An attempt will be made to find how to make machines uselanguage, form abstraction and concepts ... and improvethemselves.
John McCarthy, 1956
20
![Page 23: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/23.jpg)
Dartmouth: when AI was coined
We propose a study of artificial intelligence [..]. The study is toproceed on the basis of the conjecture that every aspect oflearning or any other feature of intelligence can in principle beso precisely described that a machine can be made to simulate it.
An attempt will be made to find how to make machines uselanguage, form abstraction and concepts ... and improvethemselves. John McCarthy, 1956
20
![Page 24: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/24.jpg)
Before AI, the vision was there:
Machine Learning, 1950by (...) mimicking education, weshould hope to modify the machineuntil it could be relied on to producedefinite reactions to certaincommands.
How ?One could carry through theorganization of an intelligentmachine with only two interferinginputs, one for pleasure or reward,and the other for pain orpunishment.
More ?http://www.csee.umbc.edu/courses/471/papers/turing.pdf
21
![Page 25: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/25.jpg)
Before AI, the vision was there:
Machine Learning, 1950by (...) mimicking education, weshould hope to modify the machineuntil it could be relied on to producedefinite reactions to certaincommands.
How ?One could carry through theorganization of an intelligentmachine with only two interferinginputs, one for pleasure or reward,and the other for pain orpunishment.
More ?http://www.csee.umbc.edu/courses/471/papers/turing.pdf
21
![Page 26: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/26.jpg)
The imitation game
The criterion:Whether the machine could answer questions in such a waythat it will be extremely difficult to guess whether theanswers are given by a man, or by the machine
Critical issueThe extent we regard something as behaving in an intelligentmanner is determined as much by our own state of mind andtraining, as by the properties of the object underconsideration.
22
![Page 27: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/27.jpg)
The imitation game, 2
A regret-like criterion
I Comparison to reference performance (oracle)
I More difficult task 6⇒ higher regret
Oracle = human being
I Social intelligence matters
I Weaknesses are OK.
23
![Page 28: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/28.jpg)
Debuts radieux. Promesses
1955 : Logic Theorist Newell, Simon, Shaw, 1955
I Relecture de Principia MathematicaWhitehead and Russell, 1910-1913 ... an attempt to derive all
mathematical truths from a well-defined set of axioms and inference rules
in symbolic logic
I General Problem Solver Newell, Shaw, Simon, 1960
Within 10 years, a computer will
I be the world’s chess champion
I prove an important theorem in maths
I compose good music
I set up the language for theoretical psychology
24
![Page 29: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/29.jpg)
Overview
Les racines : IA
IA as search
IA and games
Promesses tenues ?
What’s new
25
![Page 30: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/30.jpg)
Probleme pose
Symboles Operateurs
nombres 2 + 2 = 4
conceptsA, A→ B
|= B
Manipulation des symboles
I Nombres, operateurs arithmetiques interpretation(+,×, ...) Arithmetics, Constraint Satisfaction
I Concepts, operateurs logiquesI Propositionnel Inference, Constraint SatisfactionI Relationnel + unification
(homme(X ),mortel(X ), homme(Socrate))Logic programming
Unification + Interpretation = Constraint Programming
26
![Page 31: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/31.jpg)
Calcul symbolique, ingredients
Raisonner; parcourir un espace (arbre) de recherche
I Etats; noeuds de l’arbre
I Navigation: choix d’operateurs: transition entre etats
Comment
I Bons choix d’operateurs
I Evaluation de l’etat
I Elagage de l’arbre de recherche
Langages IPL, Lisp, Prolog
I Listes
I Actions
27
![Page 32: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/32.jpg)
Intelligence Artificielle as Search
Espace Navigation CriteresLogic +Systemes Experts + +Jeux + +
28
![Page 33: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/33.jpg)
Inference
Deduction
I Modus ponensA,A→ B
|= B
I Modus tollens¬B,A→ B
|= ¬A
Commentaire
I Truth preserving |=I Choix de la deduction
More ?http://homepages.math.uic.edu/ kauffman/Robbins.htm
29
![Page 34: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/34.jpg)
Inference, 2
Induction¬A,B
(inference) A→ B
Correlation et causalite
I Beaucoup de tuberculeux meurent a la montagne
I Donc ?
30
![Page 35: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/35.jpg)
Inference, 2
Induction¬A,B
(inference) A→ B
Correlation et causalite
I Beaucoup de tuberculeux meurent a la montagne
I Donc ?
30
![Page 36: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/36.jpg)
Inference, 3
AbductionB, A→ B
(inference) A
Causes multiples
I Si on est ivre, on titube. Or tu titubes.
I Donc ?
31
![Page 37: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/37.jpg)
Inference, 3
AbductionB, A→ B
(inference) A
Causes multiples
I Si on est ivre, on titube. Or tu titubes.
I Donc ?
31
![Page 38: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/38.jpg)
Coup d’arret
1972 : L’hiver de l’IA Rapport Dreyfus
I Une application locomotive: la traduction automatiqueI Il est necessaire de comprendre pour traduire
I Paul va passer sous le bus; Paul est un ami; je pousse PaulI Paul va passer sous le bus; Paul est un ennemi; je pousse Paul
Discussion
I Chacun sait deduire; l’expert sait raisonner
I Chambre chinoise Searle, “strong AI”
I Test de Turing
32
![Page 39: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/39.jpg)
Intelligence Artificielle as Search
Espace Navigation CriteresLogic +Systemes Experts + +
Jeux + +
33
![Page 40: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/40.jpg)
Les systemes experts
Declaratif Procedural
d’Inference
Base de
SYSTEME EXPERT
Connaissances
Moteur
Le coeur
I Base de connaissancesA→ B
I Moteur d’inferences|=, inference
34
![Page 41: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/41.jpg)
Les systemes experts
Declaratif Procedural
d’Inference
Base de
SYSTEME EXPERT
Connaissances
Moteur
Marche avant
Input Faits A
Output Diagnostic B
Marche arriere
Input B?
Output A? questions sur les faitspermettant de deduire B
35
![Page 42: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/42.jpg)
Lecons
Programmation declarative
I Non pas des ordres aux deux sens du terme
I Mais des informations
Succes
I Dendral: chimie organique Feigenbaum et al. 60s
I Mycin: medecine Shortliffe, 76
I Molgen: biologie moleculaire Stefik, 81
I R1: assemblage informatique McDermott, 82
Limites
I Rendements decroissants
I Facteurs humains
I Declaratif... mais besoin de controle
Goulet d’etranglementD’ou viennent les bases de connaissances ?
36
![Page 43: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/43.jpg)
Overview
Les racines : IA
IA as search
IA and games
Promesses tenues ?
What’s new
37
![Page 44: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/44.jpg)
Intelligence Artificielle as Search
Espace Navigation CriteresLogic +Systemes Experts + +Jeux + +
38
![Page 45: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/45.jpg)
Why games ?
I Micro-worlds finite number of states, actions
I Simple rules known transitions (no simulator needed)
I Profound complexity proof of principle of AI
39
![Page 46: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/46.jpg)
MiniMax algorithm: brute force
backward induction; Nash equilibrium
The algorithm
1. Deploy the full game tree
2. Apply utility function to terminal states
3. Backward inductionI On Max ply, assign max. payoff moveI On Min ply, assign min. payoff move
4. At root, Max selects the move with maxpayoff.
Comments
I Perfect play for deterministic, perfect information games
I Assumes perfect opponent
I Impractical: time and space complexity in O(bd)
40
![Page 47: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/47.jpg)
MiniMax algorithm: brute force
backward induction; Nash equilibrium
The algorithm
1. Deploy the full game tree
2. Apply utility function to terminal states
3. Backward inductionI On Max ply, assign max. payoff moveI On Min ply, assign min. payoff move
4. At root, Max selects the move with maxpayoff.
Comments
I Perfect play for deterministic, perfect information games
I Assumes perfect opponent
I Impractical: time and space complexity in O(bd)
40
![Page 48: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/48.jpg)
Alpha-Beta: MiniMax with pruning
alphabeta(node, depth, α, β, Player)
I if depth = 0, return H(node)
I if Player = MaxFor each child node,
α := max(α, alphabeta(child, depth-1,α, β, not(Player)))
if β ≤ α, cut beta cut-offreturn α
I if Player = MinFor each child node,
β := min(β, alphabeta(child, depth-1,α, β, not(Player)))
if β ≤ α, cut alpha cut-offreturn β
41
![Page 49: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/49.jpg)
Alpha-Beta: MiniMax with pruning
Comments
I Pruning does not affect finalresult
I Good move ordering →complexity O(b
d2 )
I Same as√
branching factorfor chess: 35→ 6.
42
![Page 50: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/50.jpg)
Chess: Deep Blue vs Kasparov
Ingredients
I Brute force; 200 millionpositions per second
I Look-ahead 12 plies
I Alpha-beta
I Tuning the heuristicfunction on a game archive
I Branching factor b ∼ 35good move ordering b ∼ 6
Controversyhttp://www.slideshare.net/toxygen/kasparov-vs-deep-blue
43
![Page 51: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/51.jpg)
Dynamic programming
Principle
I Recursively decompose the problem in subproblems
I Solve and propagate
An example
44
![Page 52: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/52.jpg)
Dynamic programming & Learning
Backgammon Gerald Tesauro, 89-95
I State: raw description of a game (number of White or Blackcheckers at each location) IRD
I Data: set of games
I A game: sequence of states x1, . . . xT ; value on last yT : winsor loses
45
![Page 53: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/53.jpg)
Dynamic programming & Learning
Learning
I Learned: F : IRD 7→ [0, 1] s.t.
Minimize |F (xT )− yT |; |F (x`)− F (x`+1)|
I Search space: F is a neural net ≡ w IRd
I Learning rule 200,000 games
∆w = α(F (x`+1)− F (x`))∑k=1
λ`−k∇wF (xk)
46
![Page 54: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/54.jpg)
Overview
Les racines : IA
IA as search
IA and games
Promesses tenues ?
What’s new
47
![Page 55: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/55.jpg)
La promesse (1960)
Within 10 years, a computer will
I be the world’s chess champion
I prove an important theorem in maths
I compose good music
I set up the language for theoretical psychology
48
![Page 56: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/56.jpg)
L’IA a beaucoup promis
The world’s chess champion ?
DiscussionEntre intelligence et force brute.
49
![Page 57: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/57.jpg)
L’IA a beaucoup promis, 2
Prouver un theoreme ?
The robot scientist
I Faits → Hypotheses
I Hypotheses → Experiences
I Experiences → Faits
I King R. D., Whelan, K. E., Jones, F. M., Reiser, P. G. K., Bryant, C. H.,Muggleton, S., Kell, D. B. and Oliver, S. G. (2004) Functional genomichypothesis generation and experimentation by a robot scientist. Nature427 (6971) p247-252
I King R.D., Rowland J., Oliver S.G, Young M., Aubrey W., Byrne E.,Liakata M., Markham M., Pir P., Soldatova L., Sparkes A., Whelan K.E.,Clare A. (2009). The Automation of Science. Science 324 (5923): 85-89,3rd April 2009
50
![Page 59: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/59.jpg)
The Concept of a Robot Scientist
Background Knowledge
Analysis
Final Theory Experiment selection Robot
Results Interpretation
Computer systems capable of originating their own experiments, physically executing them, interpreting the
results, and then repeating the cycle.
Hypothesis Formation
![Page 60: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/60.jpg)
L’IA a beaucoup promis, 3
Composer de la bonne musique ?
Musac
51
![Page 61: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/61.jpg)
L’IA a beaucoup promis, 4
Set up the language for theoretical psychology ?
Neuro-imagerie − Interfaces Cerveau-Machine
52
![Page 62: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/62.jpg)
L’IA a beaucoup promis, 4
Set up the language for theoretical psychology ?
Test d’hypotheses multipleshttp://videolectures.net/msht07 baillet mht/
53
![Page 63: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/63.jpg)
Overview
Les racines : IA
IA as search
IA and games
Promesses tenues ?
What’s new
54
![Page 64: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/64.jpg)
AI: The map and the territory
The 2005 DARPA Challenge
AI Agenda: What remains to be done Thrun 2005
I Reasoning 10%I Dialog 60%I Perception 90%
55
![Page 65: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/65.jpg)
AI: Complete agent principles
Rolf Pfeiffer, Josh Bongard, Max Lungarella,
Jurgen Schmidhuber, Luc Steels, Pierre-Yves Oudeyer...
Situated cognitionIntelligence: not a goal, a means
brains are first and foremost control systems for embodied agents,and their most important job is to help such agents flourish.
Agent’s goals: Intelligence is a means of
I Surviving
I Setting and completing self-driven tasks
I Completing prescribed tasks
What are the designer’s goals ?
56
![Page 66: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/66.jpg)
Research modes
Historical AI
I Identify sub-tasks
I Solve them
Bounded rationalityIn complex real-world situations, optimization
becomes approximate optimization since the descriptionof the real world is radically simplified until reduced to a
degree of complication that the decision maker can handle.
Satisficing seeks simplification in a somewhat differentdirection, retaining more of the detail of the real-world
situation, but settling for a satisfactory, ratherthan approximate best, decision.
Herbert Simon, 1982
57
![Page 67: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/67.jpg)
Lessons from 50 years
I We need descriptive knowledge: perceptual primitives,patterns, constraints, rules,
I We need control knowledge: policy, adaptation
I Knowledge can hardly be given: must be acquired
I We need interaction knowledge: retrieving new information,feedback
Meta-knowledge J. Pitrat, 2009
I Each goal, a new learning algorithm ?
I Problem reduction ? John Langford, http://hunch.net/
58
![Page 68: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/68.jpg)
Artificial Intelligence
Search space ML
I Representation (Un) Supervised L.
I Patterns, Rules, Constraints (knowledge)(Un) Supervised L., Data Mining
I Navigation policy Reinforcement L.
Navigation
I Inference Optimisation
Validation, control, feedback
I Criteria Statistics
59
![Page 69: L3 Apprentissagesebag/Slides/Cachan_1_2013.pdf · Exam Final: I Questions I Problems Volunteers I Some pointers are in the slides More ? here a paper or url I Volunteers: read material,](https://reader034.vdocuments.mx/reader034/viewer/2022052016/602f55cb47e0a9664c2f3f51/html5/thumbnails/69.jpg)
Questions
I Document: Perils and Promises of Big Datahttp://www.thinkbiganalytics.com/uploads/Aspen-Big Data.pdf
I Quand les donnees disponibles augmententqu’est-ce qui est different ?
I Des limitations ?
60