Vous êtes sur la page 1sur 15

Organisation

4 séances de 3h (cours + TD/TP)


Introduction
m 

m  Programme et références sur :


à http://www.agroparistech.fr/ufr-info/membres/cornuejols/Teaching/AGRO/Cours-IA/

l intelligence artificielle
iodaa-cours-IA.html

m  Contrôle continu :
❏  2 DMs
Antoine Cornuéjols ❏  Quizz (10 )
AgroParisTech ❏  Mini-Projet commun avec programmation et optimisation
antoine.cornuejols@agroparistech.fr

http://www.agroparistech.fr/ufr-info/membres/cornuejols/introduction.html

Cours IA (A. Cornuéjols) Cours IA (A. Cornuéjols) 2/57

Plan Plan du cours n°1

1- Introduc)on à l IA ; Résolu.on de problèmes par explora)on de graphes


1- Introduction à l IA

1.1- C est quoi l intelligence ?


2- Résolu.on de problèmes en environnement non déterministe (avec adversaire)
1.2- Comment étudier l intelligence ?
3- Appren)ssage par renforcement

4- Planifica)on. Formalisme STRIPS. Autres méthodes

5- Bayesian reasoning

6- Bayesian networks

7- HMM and Co

Cours IA (A. Cornuéjols) 3/57 Cours IA (A. Cornuéjols) 4/57


Exemple : jeux Exemple : robotique

➥  Perception
➥  Prise de décision
➥  Action
➥  Communication

➥  Prise de décision ➥  Apprentissage

Cours IA (A. Cornuéjols) 5/57 Cours IA (A. Cornuéjols) 6/57

Exemple : robotique Exemple : robotique (3)

m  Projet COG au MIT

m  Analyse et
interprétation
de scène

Cours IA (A. Cornuéjols) 7/57 Cours IA (A. Cornuéjols) 8/57


Exemple : Analyse des flux de polluants Exemple : « intelligence ambiante»

Projet ONU Action européenne (Kdubiq)


m  Capteurs disséminés m  Omniprésence d objets « intelligents »

m  Aspects spatial et temporel m  Percevant leur environnement

m  Analyse de documents sur m  Communiquant entre eux


le web
m  Prenant des décisions
§  Rationalité limitée
Objectifs :
Objectifs :
m 
m 
§  Bande passante limitée
Prédiction
Aide aux usagers
❏ 
❏  §  Environnements changeants
Compréhension

❏ 
❏  §  Apprentissage

Cours IA (A. Cornuéjols) 9/57 Cours IA (A. Cornuéjols) 10/57

Exemple : Systèmes ubiquitaires « intelligents » Les manifestations de la cognition

Applications
m  Discrimination
m  Domotique
m  Adaptation
m  Aide aux personnes handicapées
m  Perception / associations / analogies / métaphores
m  Station orbitale

m  … m  Communication

m  Planification

m  Apprentissage et découverte

Cours IA (A. Cornuéjols) 11/57 Cours IA (A. Cornuéjols) 12/57


Qu est-ce que l IA ? Qu est-ce que l intelligence artificielle ?

« The study of the conjecture that every aspect of learning or

any other feature of intelligence can in principle be so precisely

described that a machine can be made to simulate it. »

Dartmouth Workshop, Summer of 1956

Cours IA (A. Cornuéjols) 13/57 Cours IA (A. Cornuéjols) 14/57

Objectifs de l IA ? Sources de difficultés ?

m  Construire des machines pour une large variété d applications m  La perception est ambiguë de manière inhérente
augmentant les capacités de résolution de problèmes « mal posés »
❏  « J ai laissé tomber l œuf sur le sol et il s est cassé »
m  Formaliser ce qu est la connaissance
m  Rationalité limitée
m  Mécaniser l intelligence
❏  Y-a-t-il un premier coup gagnant aux échecs ?

m  Utiliser des modèles computationnels pour comprendre des


comportements complexes m  Compromis entre buts en conflits

m  Rendre l interaction avec les systèmes computationnels m  Information incomplète et/ou incertaine
aussi simple qu avec les humains

Cours IA (A. Cornuéjols) 15/57 Cours IA (A. Cornuéjols) 16/57


L IA par rapport à l informatique Caractéristiques de l IA

❏  Etudie les problèmes insolubles en informatique classique m  Prise en compte de la Connaissance (symbolique)

m  Solution optimale inconnue ou trop difficile à trouver/identifier


m  Raisonnements
m  Des paramètres difficiles à maîtriser : connaissances incertaines, vagues, ❏  Logiques / Probabilistes / imprécis / …

hypothétiques, ... ❏  Induction / analogie / …

❏  Heuristique

❏  Exigences d explication ou de communication sophistiquée avec ❏  Non déterministe vs. déterministe

les utilisateurs
m  Solutions satisfaisantes vs. Optimales

m  Pas programmé mais par apprentissage


❏  Objectif d adaptation et d apprentissage avec l expérience

...
❏  IA (A. Cornuéjols)
Cours 17/57 Cours IA (A. Cornuéjols) 18/57

Domaines d applications de l IA Domaines d applications de l IA

m  Ordonnancement et planifica0on
❏  Système DART de la DARPA u0lisé pour la logis&que des opéra0ons « Desert storm » et m  Systèmes de diagnostic
« Desert shield »
❏  Système Pathfinder et Internist en médecine (supérieurs aux médecins)
❏  Système de changement de trajets pour American Airlines
❏  Système d aide aux consommateurs (Whirlpool)
❏  Système de planifica&on et d ordonnancement pour l assemblage des vols spa0aux
européens (ESA)
m  Systèmes de conception et configuration

m  Reconnaissance de la parole ❏  Système XCON de DEC pour aider à la configuration de systèmes


❏  Émergence d’assistants sur les smartphones informatiques

m  Vision par ordinateur m  Aide à la décision (e.g. financière)


❏  Systèmes de reconnaissance de visages, systèmes de surveillance (y compris suivi entre
❏  Systèmes de détection de fraude (IRS aux USA, banques (cartes de crédit))
caméras)
❏  Aide aux demandes de financement des particuliers
❏  DARPA « Grand Challenge » et DARPA « Urban Challenge »
❏  Reconnaissance de l écriture manuscrite

Cours IA (A. Cornuéjols) 19/57 Cours IA (A. Cornuéjols) 20/57


Domaines d applications de l IA Les approches de l IA

m  Systèmes de classifica0on m  L école symbolique


❏  SKYSURVEY de la NASA ❏  La pensée consiste en une séquence de transformations de représentations

➥  Agents cognitifs délibératifs


m  Démonstra0on automa0que de théorèmes
m  L école sub-symbolique
m  Systèmes de jeu ❏  La pensée consiste en une transmission et une transformation de signaux
Champions du monde aux dames (anglo-saxonnes), othello, backgammon, échecs, go
❏ 
➥  Agents réactifs
❏  Le pocker ? Le bridge : pas encore

❏  Sert à concevoir des systèmes s adaptant à des milieux hos0les m  L école de la cognition située
❏  La pensée ne réside pas dans la tête d un individu isolée mais est distribuée
m  Traduc0on automa0que au sein d une société (incluant son histoire et sa culture) et suscitée par le
❏  Systèmes de traduc0on de Google (bientôt en temps réel) (MAIS …) contexte.

➥  Emergence et auto-organisation
Cours IA (A. Cornuéjols) 21/57 Cours IA (A. Cornuéjols) 22/57

Bref historique (1) Biological Neurons

0. Depuis l Antiquité ...


❏  Bacon / Pascal / Descartes / Leibniz
❏  Babbage & Ada de Lovelace / Boole

❏  Réflexion sur les fondements des mathématiques

1. Les pionniers de l IA (~1936 - ~1956)


❏  Turing : les fonctions calculables sont définies par une machine

❏  Von Neumann : architecture des ordinateurs, réflexions sur les automates


McCulloch & Pitts (1943) : Premier modèle formel du neurone Terminal Branches
❏  Dendrites
of Axon
❏  Hebb (1949) : assemblées de neurones et règle d apprentissage
❏  Les conférences Macy et la première Cybernétique

Axon

Cours IA (A. Cornuéjols) 23/57 Cours IA (A. Cornuéjols) 24/57


1er modèle formel de neurone Bref historique (2)

2. L IA comme méthodes générales (~1956 - ~1968)


Terminal Branches
Dendrites La pensée comme manipulation de représentations discrètes des
of Axon ❏ 

x1
w1
connaissances: IA symbolique
x2
w2 m  Démonstrateurs de théorèmes (principe de résolution (-> Prolog))
x3 w3
m  Résolveurs universels de problèmes (GPS)
S
m  Le système CHECKER
Axon

wn
❏  Premier connexionnisme
xn
m  Le Perceptron [Rosenblatt, 1957-1962]

❏  MAIS des échecs :


McCulloch & Pitts (1943) m  La traduc0on automa0que
m  Les experts ne sont pas des experts universels !?
m  Le Perceptron est limité
Cours IA (A. Cornuéjols) 25/57 Cours IA (A. Cornuéjols) 26/57

L’espoir Raisonnement / résolution de problèmes

État
A B C
((A)(B)(C))
Opérateur
L’intelligence met en jeu

des processus généraux de raisonnement

boug
,A
er(B

er(B
boug

,C
A

)
BC C
A A B
BC C
((AB) (C)) B B A B ((A) (CB))
A C A C
((B) (AC)) ((CA) (B))
((BA) (C)) ((A) (BC))

Cours IA (A. Cornuéjols) 27/57 Cours IA (A. Cornuéjols) 28/57


5 Early Heuristic Programs

5.1
Raisonnement / résolution de problèmes The Logic Theorist andgoal
None of these represents the goal configuration, so we continue to apply
transformations to each of these and so on until a structure representing the
is reached.
Heuristic SearchWe (and the computer) can keep track of the transformations
Un exemple d'induction : inversion de la résolution
made by arranging them in a treelike structure such as shown in Fig. 5.3.
(The arrowheads on both ends of the lines representing the transformations
indicate that each transformation is reversible.) +1
-1

Figure 5.2: The eight-puzzle.

with. I will represent the starting position by the following structure, which is
a list of three sublists:
fille(claude,jacques)
parent(X,claude)
:= fille(claude,X) fille(claude,jacques)
?
((2, 8, 3), (1, 6, 4), (7, B, 5)).

The first sublist, namely, (2, 8, 3), names the occupants of the first row of the
puzzle array, and so on. B stands for the empty cell in the middle of the third
row.
In the same fashion, the goal configuration is represented by the following
structure:
((1, 2, 3), (8, B, 4), (7, 6, 5)).

Next, we have to show how a computer can transform structures of the


kind we have set up in a way that corresponds to the allowed moves of the
puzzle. Note that when a tile is moved, it swaps places with the blank cell;
that is, the blank cell moves too. The blank cell can either move within its row
or can change rows.
Corresponding to these moves of the blank cell, when a tile moves within parent(jacques,claude)
its row, B swaps places with the number either to its left in its list (if there is parent(jacques,claude)
m  Recherche dans un graphe
one) or to its right (if there is one). A computer can easily make either of
these transformations. When the blank cell moves up or down, B swaps places
Figure 5.3: A search tree.

with the number in the corresponding position in the list to the leftThis version
(if there is of the eight-puzzle is relatively simple, so not many
one) or in the list to the right (if there is one). These transformations can also have to be tried before the goal is reached. Typically though
transformations
be made quite easily by a computer program. (especially in larger versions of the puzzle), the computer would be swamped
Cours IA (A. Cornuéjols) 29/57 Cours IA (A. Cornuéjols) 30/57
Using the Newell and Simon approach, we start with theby all ofstructure
symbol the possible transformations – so much so that it would never
generate
representing the starting configuration of the eight-puzzle and a goal expression. To constrain what was later called “the
apply allowed
combinatorial
transformations until a goal is reached. There are three transformations explosion” of transformations, Newell and Simon suggested
of the
starting symbol structure. These produce the following structures:
using “heuristics” to generate only those transformations guessed as likely to
be on the path to a solution.
((2, 8, 3), (1, 6, 4), (B, 7, 5)),
In one of their papers about LT, they wrote “A process that may solve a
((2, 8, 3), (1, 6, 4), (7, 5, B)), problem, but o↵ers no guarantees of doing so, is called a heuristic for that
and problem.” Rather than blindly striking out in all directions in a search for a
((2, 8, 3), (1, B, 4), (7, 6, 5)).

Premier connexionnisme : le perceptron


Copyright c 2010 Nils J. Nilsson
116
115
Copyright c 2010 Nils J. Nilsson
The perceptron: a linear discriminant
http://ai.stanford.edu/⇠nilsson/
http://ai.stanford.edu/⇠nilsson/ All rights reserved. Please do not reproduce or cite this version. September 13, 2009.
All rights reserved. Please do not reproduce or cite this version. September 13, 2009. Print version published by Cambridge University Press.
Print version published by Cambridge University Press.
http://www.cambridge.org/us/0521122937
http://www.cambridge.org/us/0521122937

m  Frank Rosenbla; (1958 – 1962)

!
n
Ψ(x) = wi φi (x)
i=1

Cours IA (A. Cornuéjols) 31/57 Cours IA (A. Cornuéjols) 32/57


6 Semantic Representations

1 9 6 8 : que sait-on faire ? intelligence-test questions. Each member of


We shall be considering the solution by machine of so-called
“geometric-analogy” 1 9 6 8 : que sait-on faire ? 4 Pattern Recognition
this class of problems consists of a set of labeled line drawings. The
task to be performed can be described by the question: “Figure A
is to Figure B as Figure C is to which of the following figures?” For
example [in Fig. 6.1] it seems safe to say that most people would
m  Raisonnement shown in Fig. 4.10.

m  Raisonnement
agree with the program we are about to describe, in choosing
[number 4] as the desired answer. ❏  Résolution de problème
❏  Résolution de problème ❏  Mémoire associative
❏  Mémoire associative
m  Jeu de dames
m  Jeu de dames
m  Analogie m  Analogie

m  Conversation (?) m  Conversation (?)


❏  Eliza ❏  Eliza
Apprentissage
Apprentissage
m 
m 

❏  Reconnaissance de caractères Figure 6.1: An analogy problem.


❏  Reconnaissance de Figure 4.10: Recognition of FORTRAN characters. Input is above and output
❏  Jeu de dames He further noted that “problems of this type are widely regarded as (with only two errors) is below. (Illustration used with permission of SRI Inter-
Taux de reconnaissance = 98% !!!
requiring a high degree of intelligence for their solution and in fact are used as caractères national.)
a touchstone of intelligence in some general intelligence tests used for college
admission and other purposes.” So, again, AI research concentrated on After the neural net part of the system was trained, the overall system
mechanizing tasks requiring human intelligence. ❏  Jeu de dames (which decided on the most confident legal string) was able to achieve a
recognition accuracy of just over 98% on a large sample of material that was
Evans’s program first transformed the diagrams presented to it so that not part of what the system was trained on. Recognizing handwritten
they revealed how they were composed out of parts. He called these characters with this level of accuracy was a significant achievement in the
Cours IA (A. Cornuéjols) 33/57
“articular” representations. Of the possibly several decompositions possible, Cours IA (A. Cornuéjols) 34/57
1960s.18
the one chosen by the program depended on its “context.” (This choice is one
Expanding its interests beyond neural networks, the Learning Machines
example of a heuristic used by the program.) For example, the diagram
Group ultimately became the SRI Artificial Intelligence Center, which
continues today as a leading AI research enterprise.

4.3 Statistical Methods


9.3 Computer Vision of Three-Dimensional Solid Objects
During the 1950s and 1960s there were several applications of statistical

Vision artificielle Robotique mobile methods to pattern-recognition problems. Many of these methods bore a close
12.1resemblance to some of the neural network techniques.
Shakey, Recall that earlier I
the SRI Robot
explained how to decide which of two tones was present in a noisy radio signal.
could either be decomposed into A similar technique could be used for pattern recognition. For classifying
m  Stanford AI Lab 132
9.3 Computer Vision of Three-Dimensional Solid Objects images (or other perceptual inputs), it was usual to represent the input by a

8 1960s’ Infrastructure 102


Max Clowesc (circa
Copyright 2010 Nils J. Nilsson
1944–1981) of Sussex University in Britain developed
Vision + planification +
http://ai.stanford.edu/⇠27nilsson/
similar Please
All rights reserved. ideas independently,
do not reproduceand the this
or cite labeling scheme
version. is now 13,
September generally
2009. known Copyright c 2010 Nils J. Nilsson
support from ARPA, the Lab took delivery of a DEC PDP-6 computer and,
later, a PDP-10 computer. In addition to its work in AI (which I’ll describe in as Hu↵man–Clowes
Print version published labeling.
by Cambridge University Press. http://ai.stanford.edu/⇠nilsson/
All rights reserved. Please do not reproduce or cite this version. September 13, 2009.

interface par langage


subsequent chapters), SAIL was involved in many other computer-related http://www.cambridge.org/us/0521122937
Next comes David Waltz (1943– ). In his 1972 MIT Ph.D. thesis, he Print version published by Cambridge University Press.
projects including the development of a precursor to computer “windows” and
the early installation of terminals in everyone’s offices.5 extended the Hu↵man–Clowes line-labeling scheme to allow for line drawings http://www.cambridge.org/us/0521122937
of scenes with shadows and possible “cracks” between two adjoining objects.28
Waltz’s important contribution was to propose and implement an efficient
computational method for satisfying the constraint that all of the lines must pseudo naturel
be assigned one and only one label. (For example, an edge can’t be concave
at one end and convex at the other.) In Fig. 9.15, I show an example of a line
drawing that Waltz’s program could correctly segment into its constituents.

Figure 9.12: A scene analyzed by SEE. (Illustration used with permission of


Adolpho Guzman.)

leaving MIT in 1967 to become a professor of Information and Computer


Science at the University of California at Santa Cruz, he completed a theory
for assigning labels to the lines in drawings of trihedral solids – objects in
which exactly three planar surfaces join at each vertex of the object. The
Figure 8.1: Site of the Stanford AI Lab from 1966 until 1980. (Photograph labels depended on the ways in which planes could come together at a vertex.
courtesy of Lester Earnest.) (I got to know Hu↵man well at that time because he consulted frequently at
the Stanford Research Institute.)
Since their early days, the groups at CMU, MIT, and Stanford have been
among the leaders of research in AI. Often graduates of one of these Hu↵man pointed out that there are only four ways in which three plane
institutions became faculty members of one of the other ones.
Figuresurfaces
9.15: Acan come
scene withtogether
shadows at analyzed
a vertex.25 byThese
Waltz’s areprogram.
shown in(Illustration
Fig. 9.13. In
Figure 12.3: Shakey as it existed in November 1968 (with some of its components
Around 1965 another world-class AI center emerged at the University of addition
used with to theseoffour
permission Davidkinds of vertices, a scene might contain what Hu↵man
Waltz.) labeled). (Photograph courtesy of SRI International.)
Edinburgh in Scotland. Its founder was Donald Michie (1923–2007; Fig. 8.2), called “T-nodes” – line intersection types caused by one object in a scene
who had worked with Alan Turing and I. J. (Jack) Good at Bletchley Park
during the Second World War. Discussions there with Turing and Good about occluding another.
Summarizing some ofThese all give
the work rise to a number
on processing of di↵erent
line drawings kinds of labels
at MIT,
forWinston
the linessays
in the scene; these was
labels around a single obstacle lying between its initial position and a goal position,
intelligent machines captivated Michie. As he reported in an October 2002 Patrick that “Guzman thespecify whether theHu↵man
experimentalist, lines correspond
the to
interview, “I resolved to make machine intelligence my life as soon as such an convex, concave, or occluding edges. Shakey should first head toward a point near an occluding boundary of the
theoretician,
Cours IAbecame
enterprise (A. feasible.”
Cornuéjols)
6
Because computer facilities in the mid- to late 35/57 and Waltz the encyclopedist (because Waltz had to catalog Cours IA (A. Cornuéjols) 36/57
obstacle and then head straight for the unobstructed final goal point.
1940s were primitive and scarce, Michie became a geneticist and molecular thousands Hu↵man
of junctions,
notedin that
orderthe
to labels
deal with
of thecracks
linesandin ashadows).”
29
drawing might be locally However, the situation becomes more complicated if the environment is
biologist. consistent (around someforvertices)
Meanwhile, similar work finding,but still be globally
identifying, inconsistent
and describing (around
objects in all littered with several obstacles, and we sought a general solution to this more
of the vertices). Consider, for example, Roger Penrose’s difficult problem.
Pursuing his interest in machine intelligence, from the sidelines as it were, three-dimensional scenes was being done at Stanford. By 1972famous line drawing of
Electrical
in 1960 he put together a “contraption of matchboxes and glass beads” that an “impossible object”Gilbert
shownFalkin Fig. 9.14. (It isscenes
impossible
of linebecause no
26
Engineering Ph.D. student could segment drawings Shakey kept information about the location of obstacles and about its
could learn to play tic-tac-toe (noughts and crosses). He named his “machine”
MENACE, an acronym for Matchbox Educable Noughts and Crosses Engine.7 three-dimensional
into separate object,
objects using viewed that
techniques in “general position,”ofcould
were extensions thoseproduce
of this
image.)
Guzman. 30
AndNo by
“real scene”
1973, can have
Computer a linePh.D.
Science with two di↵erent
student Gunnarlabels.
Grape
217
158 performed segmentation of scenes containing parallelepipeds and wedges using
them about. Unlike the the world of real blocks used in previous MIT and
Stanford hand–eye research, Winograd’s blocks world was simulated on a DEC
PDP-10 computer and viewed on a DEC 340 black-and-white graphics display
such as that shown in Fig. 13.2. SHRDLU could be given commands, typed in
ordinary English, to manipulate the blocks using its gripper. It could also be
asked questions about the state of its world. It could “understand” those
commands and questions, carry out requested manipulations, and ask the user
clarifying questions in English about questions or commands it did not
SHRLDU
understand.

Interface en langage pseudo naturel Les réseaux sémantiques


Before describing something about SHRDLU’s design, let us look at an
example of part of a dialog.8 In the following, the user’s input is numbered
and shown in lowercase letters, and SHRDLU’s response is in all caps,
Winograd’s
Figure commentary
13.2: SHRDLU follows
’s world. in the same
(Illustration typeface,
used and mineofisTerry
with permission in Wino- 13 Progress in Natural Language Processing
parentheses in normal type:
grad.)

1. pick up a big red block.


OK.
FigureIn13.2: SHRDLU
order
grad.) off system
The
’s world.
to pick
answers
(Illustration
up the
"OK" for
red block,
whenthe
used with
it green
permission
it had
carries
to clearof Terry
it Wino- [Ross Quillian, 1968 : Semantic memory]
by finding a space oneout
anda moving
command.the
green one away.
(I won’t bother to show what the block configuration looks like as the dialog
OK. assuming you can use your own mental “model” to keep track of
proceeds, 239
things.The
Thesystem
gripper is now holding
answers the red
"OK" when itblock.)
carries out a command.
Copyright c 2010 Nils J. Nilsson
http://ai.stanford.edu/⇠nilsson/
All rights reserved. Please do not reproduce or cite this version. September 13, 2009.
m  Résolution d'ambiguïtés
2. bother
(I won’t grasp
Print tothe
showpyramid.
what the block configuration looks like as the dialog
proceeds,
Thereassuming
version
are three
published by Cambridge
use your
pyramids in own
University Press.
youhttp://www.cambridge.org/us/0521122937
can themental
scene,“model”
so thetoprogram
keep track of (par intersection de propagation de marqueurs)
things. Thenothing.
does gripper is now holding the red block.)
13.2 Understanding
I DON’T UNDERSTAND WHICH PYRAMID YOU MEAN.
2. grasp the pyramid. "L'astronome voulait épouser une étoile"
3. find a block which is taller than the one you are
(The user has are
There
holding now apparently
andthree
put it changed
pyramids
into inhis
the mind
the
box. aboutsopyramid
scene, grasping.)
the program
does nothing.
240 The systemUNDERSTAND
I DON’T has to figure WHICHout what is
PYRAMID YOUmeant
MEAN.by "one" and
"it,"
3. findas well as using
aCopyright
block deduction to decide which block
cwhich isJ. taller
Nilsson than the one you are
Un concept acquiert un sens à travers le réseau sémantique dans lequel il
2010 Nils Figure 13.2: SHRDLU’s world. (Illustration used with permission of Terry Wino-
itholding
is holding.
and put it into the box.
http://ai.stanford.edu/ ⇠nilsson/
grad.) m 
All rights reserved. Please do not reproduce or cite this version. September 13, 2009.
BY "IT," I ASSUME YOU MEAN THE BLOCK WHICH IS TALLER
Print version published by Cambridge University Press.
240 THAN THE ONE I AM HOLDING. s'insère et les relations qu'il a avec d'autres concepts
http://www.cambridge.org/us/0521122937 OK.
The system answers "OK" when it carries out a command.
Printed out Copyright
to keepc 2010
us informed.
Nils J. Nilsson

The hand hashttp://ai.stanford.edu/
to set down thenilsson/
red block to pick up the
All rights reserved. Please do not reproduce or cite this version. September 13, 2009.
(I won’t bother to show what the block configuration looks like as the dialog
proceeds, assuming you can use your own mental “model” to keep track of
tallerPrintblue one.published by Cambridge University Press.
version things. The gripper is now holding the red block.)
http://www.cambridge.org/us/0521122937
OK. 2. grasp the pyramid.
4. what does the box contain? There are three pyramids in the scene, so the program
does nothing.
THE BLUE PYRAMID AND THE BLUE BLOCK. I DON’T UNDERSTAND WHICH PYRAMID YOU MEAN.
5. what is the pyramid supported by? 3. find a block which is taller than the one you are
Cours IA (A. Cornuéjols) 37/57 holding and put it into the box. Cours IA (A. Cornuéjols) 38/57
Earlier the system responded to "the pyramid" by saying
"I don’t understand which pyramid you mean." now, it 240

knows that we must mean the one which was just Copyright c 2010 Nils J. Nilsson
http://ai.stanford.edu/⇠nilsson/
All rights reserved. Please do not reproduce or cite this version. September 13, 2009.
mentioned. Print version published by Cambridge University Press.
http://www.cambridge.org/us/0521122937
THE BOX

This dialog illustrating SHRDLU’s capabilities goes on for forty-two more user
inputs and system responses.9
To engage in dialogs like this, SHRDLU needed “knowledge.” It needed to

Les réseaux sémantiques


know about the syntax of that part of English it used, and it needed to know
what various words and word combinations mean in the context of their use.
It also needed to know about its blocks world – how blocks can be
Les réseaux sémantiques
manipulated and what it means for an object to be inside of the “box.” It
needed to keep track of the dialog so that it could decide to which object
"L'astronome veut épouser une étoile"
mentioned previously a word such as “it” referred.
All of this needed knowledge was represented in LISP programs, or
“procedures,” as Winograd called them. Knowledge about syntax was Sorte-de
represented as a collection of procedures based on the principles of “systemic Homme
Astronome
grammar.”10 Knowledge about the meanings of words in context was Epouser Paul
represented in procedures that could refer to a dictionary of word meanings, to Propriétaire
other parts of the sentence in which the word was used, and to the discourse.
Knowledge about the blocks world was represented in two ways: There was a
model that gave the locations of all of the objects and there were procedures Sorte-de
that could infer the predicted e↵ects (in the model) of manipulations by the Possédé Titine Voiture
gripper on the various objects. The object-moving procedures had information Personne
Astre
both about the preconditions and about the e↵ects of these manipulations. Possession-1
These procedures were encoded in a version of Hewitt’s PLANNER language, Début
Profession
which, as mentioned previously, bore some resemblance to STRIPS operators.
Célébrité Sorte-de
241 Fin
Printemps-86 Printemps
Planètes
Copyright c 2010 Nils J. Nilsson
http://ai.stanford.edu/⇠nilsson/
Etoile
All rights reserved. Please do not reproduce or cite this version. September 13, 2009. Sorte-de
Print version published by Cambridge University Press.
http://www.cambridge.org/us/0521122937
Sorte-de
Observatoire Soleil Automne-87 Automne
Cinéma
Danseuse
Sorte-de
Possession Situation
Inférence par propagation d'activité ou de marqueurs : focalise l'attention
Cours IA (A. Cornuéjols) 39/57 Cours IA (A. Cornuéjols) 40/57
Bref historique (3) Perspective historique : apprentissage automatique
m  Expériences de pensée sous contrainte de réalisabilité
1950s computationnelle
Opérateurs sur des représentation
3. Knowledge is power (~1968 - ~1980) § 

§  Recherche dans un espace d’états. Buts / sous-buts


Apprentissage par mutations aléatoires (mais guidée par ressemblance) ~ AG
Représentation des connaissances
§ 
❏ 
§  Ivresse : comprendre la pensée
m  Représentations structurées : réseaux sémantiques, scripts, schémas et frames §  « Our problem, our joint problem, is to discover what transformations must be
made on the available data in order to preserve intact the significant features
(précurseurs des LO), ...
and to discard the irrelevant details ».
m  Des extensions de la logique : logiques non monotones, logiques temporelles, logique
1960s m  Principes, théorèmes et démonstrations (Checker. Problèmes
floue, ...
« jouets »)
m  De nombreux systèmes intelligents : ARCH, AM, MAGGIE, BORIS, PLANNER, ... §  Reconnaissance des formes. Plutôt numérique (bayésien, perceptron)
m  Projet CYC 1970s
❏  Utilisation des connaissances : les systèmes experts m  Expertise. Sciences cognitives : plausibilité. Intégration dans
le raisonnement
§  Modèles de mémoire. Réseaux sémantiques. Représentation des connaissances
➥  Problème : comment acquérir toutes les connaissances nécessaires ? §  Règles de production. Moteur d’inférence.
§  Mécanismes d’apprentissage et de généralisation
Comment généraliser ces expériences (souvent appliquées à des problèmes jouet ) ?
§  Apprentissage et raisonnement : heuristiques, macro-opérateurs, chunking

Cours IA (A. Cornuéjols) 41/57 Cours IA (A. Cornuéjols) 42/57

Les systèmes experts Les systèmes experts : DENDRAL


m  Le système DENDRAL
m  Transparents SE ❏  Pour la NASA : 1965 - …

m  BORIS ❏  Y a-t-il de la vie sur Mars ?


Spectrographie de masse
Les limites
❏ 
m 

intensité

Formule développée
du composé
chimique ?

Cours IA (A. Cornuéjols) 43/57 Cours IA (A. Cornuéjols) masse


44/57
Les systèmes experts : DENDRAL Les systèmes experts : MYCIN
m  Exemples de connaissances MYCIN : Mes recommandations thérapeutiques seront basées sur les infections
❏  Règle : probables et les organismes infectieux potentiels suivant :
Si le spectre de la molécule présente deux pics x1 et x2 tels que : L organisme peut être : 1° pseudomonia aeruginosa, 2° klebsiellia
1.  x1 - x2 = M + 28 pneumoniae, 3° escherichia coli, 4° bacteroides fragilis, 5°
enterobactoriaceae, 6° proteus non mirabilis
2.  x1 - 28 est un pic élevé
3.  x2 - 28 est un pic élevé Voulez-vous me donner la gravité de la maladie de M. Fred Smith sur
4.  au moins l un des pics x1 et x2 est élevé une échelle de 0 à 4 ?
Alors la molécule contient un groupe cétone Médecin : 3
MYCIN : M. Fred Smith présente-t-il des réactions allergiques à un quelconque
R1 R1 (x1) R1 agent microbien ?
Médecin : Non
C C
Se décompose O C ou en : O C
en : MYCIN : Quel est son poids en kilos ?
R2 R2 R2 (x2)
Médecin : 70
Cours IA (A. Cornuéjols) 45/57 Cours IA (A. Cornuéjols) 46/57

10 Cours Systèmes Experts


Les systèmes experts : MYCIN Les explications des SEs
Exemple 2 : Explication par le système de son propre raisonnement

MYCIN : Ma recommandation thérapeutique sera en premier choix : pour couvrir


les points <1 2 3 5 6>, prescrire de la Gentamicine, dose 119 mg 98 h MYCIN : L'infection du malade a t-elle été contractée à l'hôpital ?

IV, pendant 10 jours; Utilisateur : pourquoi ?

TEIRESIAS : [pourquoi est-ce important de savoir cela ?]


Pour couvrir le point 4, prescrire de la Clindamycine, dose 595 mg 96 h
Cela aidera à déterminer la catégorie à laquelle appartient l'organisme responsable
IV, pendant 14 jours. de l'infection :

Désirez-vous éventuellement éliminer certains médicaments ? - il a déjà été établi que 1° l'organisme est à Gram négatif, 2° c'est un bâtonnet, 3° il
se développe à l'air ou sans air,
Médecin : Oui, la Clindamycine. - donc si l'infection par cet organisme n'a pas été contactée lors du séjour à l'hôpital

alors il est presque certain (0,8) que l'organisme est de la catégorie des
MYCIN : Pour couvrir les points <2 3 4 5 6> donner du Chloramphénicol, dose 1 g
enterobactoriaceae; de plus, il se peut (0,1 ; règle 37) que ce soit la pseudomonia
96 h IV, pendant 14 jours; aeruginosa.

** Surveiller le taux de globules blancs du patient. Utilisateur : pourquoi n'avez-vous pas envisagé la possibilité que l'organisme soit un
streptocoque ?
Pour couvrir le point 1, donner de la gentamicine, dose 119 mg 98 h IV, TEIRESIAS : la règle 33 aurait du être utilisée pour identifier l'organisme comme un
pendant 10 jours. streptocoque, mais la seconde clause "la morphologie est de type coque" était déjà
connue comme fausse et la règle 33 n'a jamais été essayée.
** Modifier la dose en cas de faiblesse rhénale. .

Cours IA (A. Cornuéjols) 47/57 Cours IA (A. Cornuéjols) 48/57

1.5 Structure des systèmes experts


Un système de production peut être vu comme constitué de trois composants de base : une base de connaissance
BORIS [Dyer,82] BORIS [Dyer,82]
"Richard hadn't heard from his college roomate Paul for years. Richard had
"Richard hadn't heard from his college roomate Paul for years. Richard had
borrowed money from Paul which was never paid back. But now he had no borrowed money from Paul which was never paid back. But now he had no
idea where to find his old friend. When a letter finally arrived from San idea where to find his old friend. When a letter finally arrived from San
Francisco, Richard was anxious to find out how Paul was." Francisco, Richard was anxious to find out how Paul was."

Comment traiter un non événement ?

"John walked in the room and Mary was not there" (John désirait voir Mary et s'attendait
à la trouver dans la pièce -> pourquoi n'y est-elle pas ? …)

Ici, on est prêt à comprendre que Richard se sent une obligation vis-à-vis de Paul. Il faut
donc des connaissances sur les relations sociales et sur la psychologie.

Comment traiter "had" (possession ?) et "old" (âge de son ami ?. Non. Mais pourquoi et
comment?)
Comment infère-t-on que cette lettre vient de Paul alors que ce n'est pas dit explicitement ?

Cours IA (A. Cornuéjols) 49/57 Cours IA (A. Cornuéjols) 50/57

Pourtant … L’espoir (suite de la suite)

m  Ingénierie des connaissances

m  Processus très lourd


L’intelligence met en jeu beaucoup de connaissances
m  Peu systématisé
que l’on obtiendra par des processus généraux d’apprentissage
m  Maintenance difficile

Dépasse le quantum d’action = la thèse

Cours IA (A. Cornuéjols) 51/57 Cours IA (A. Cornuéjols) 52/57


Bref historique (4) Bref historique (5)

4. Méthodes générales d apprentissage (~1980 - ~ now) 5. Apprentissage statistique et applications (~1990 - ...)
❏  Nouvelles méthodes subsymboliques (réseaux de neurones, algorithmes génétiques, réseaux
❏  Méthodes symboliques bayésiens, chaînes de Markov, ...)

m  De nombreuses réalisations et nouvelles techniques ❏  Théorie de l apprentissage statistique (données i.i.d.)

n  Algorithme d élimination ❏  Besoins industriels


n  Arbres de décision
n  Méthode de l Etoile ➥  On ne parle plus de connaissances

m  Nouveaux principes m  Perspectives :


n  Espace des versions ❏  Data mining ; Text mining ; Web mining
❏  BIGGG Data
❏  Renouveau du connexionnisme et du mouvement subsymbolique
❏  Systèmes à longue durée de vie
m  Hopfield (1982)
m  Le Perceptron Multi-Couches (1985) ➥  Retour de la connaissance ?
IA distribuée (Algorithmes Génétiques, Vie Artificielle, ...)
m 
➥  Vers ra.onalité limitée et systèmes complexes hétérogènes
à longue vie
Cours IA (A. Cornuéjols) 53/57 Cours IA (A. Cornuéjols) 54/57

État de l art Faire de l intelligence artificielle

Lesquelles de ces choses sont actuellement possibles ?


m  Point de vue des sciences cognitives
❏  Jouer décemment au ping-pong
❏  Comprendre la cognition naturelle par des modèles informatiques
❏  Conduire un véhicule le long d une route sinueuse en montagne

❏  Conduire un véhicule dans le centre d une grande ville m  Point de vue de l ingénieur
Jouer décemment au bridge ; Jouer décemment au pocker
Résoudre des problèmes difficiles par n importe quelle méthode
❏ 
❏ 

❏  Découvrir et prouver un nouveau théorème de mathéma0que

❏  Écrire une histoire inten0onnellement drôle m  Point de vue du théoricien


❏  Donner un conseil compétent dans un domaine précis du droit ❏  Découvrir les lois universelles de la cognition
❏  Traduire une conversa0on entre un Français et un Anglais en temps réel (dans les deux sens)

➥  Oscilla0on entre ces trois pôles

Cours IA (A. Cornuéjols) 55/57 Cours IA (A. Cornuéjols) 56/57


Ouvrages conseillés

m  R. Callan : Artificial Intelligence. Palgrave MacMillan, 2003.

m  I. Millington : Artificial Intelligence for Games. Morgan Kaufmann, 2006.

m  Nilsson N. (98) : Artificial Intelligence : A new synthesis. Morgan Kaufmann,


1998.

m  E. Rich & K. Knight : Artificial Intelligence. McGraw-Hill, 1991

m  S. Russell & P. Norvig : Artificial Intelligence: A modern approach (3nd ed.).


Prentice Hall, 2009
(Trad. française : « Intelligence artificielle », Pearson Education, 2006, 1184 p.)

Cours IA (A. Cornuéjols) 57/57

Vous aimerez peut-être aussi