Sei sulla pagina 1di 17

06.01.2019 Art f c al Intell gence.

html

Ank Glossary: Art f c al Intell gence (190 cards) Created on 2019-01-06

1. Env ronment 7. Learn ng


Types ? Agents?

2. Agentt Types?

8. what s the 4 State space


element of Act ons
search Goal test
problem? Path cost

3. S mple reflex 9. How to evaluate Completeness - does t always


agent search f nd a solut on f ex sts
strateg es? T me complex ty - number of
nodes generated/expanded
Space complex ty - max mum
number of nodes n memory
Opt mal ty - does t always f nd a
least-cost solut on

10. Un nformed Bread-f rst


4. Reflex Agent ? search Un form-cost
strateg es? Depth-f rst
Depth-l m ted
Iterat ve deep n ng

11. What s expand the start node, expand the


Breadth-f rst f rst level then the second level
search

12. Propert es of Complete - Yes( f branch ng factor


5. Goal-based
breadth-f rst s f n te)
agents? search T me - 1 + b + b2+ ... + bd + b(bd-
1)= O(bd+1)
Space - O(bd+1) keeps every node
n memory
Opt mal - Not opt mal n general
but yes f cost=1 per step
Goal test s appl ed when the
nodes are selected for expans on.
If goal test s appl ed when the
6. Advantages of -Not the only way of act ng rat onal nodes are generated the t me and
ut l ty-based -However, advantages n terms of
space complex ty O(bd)
agents flex b l ty and learn ng
-Can act rat onally n two
mportant cases where the others 13. Un form Cost -expands the least-cost node
fa l: Search ? -goal test appl ed when node s
hav ng confl ct ng goals selected for expans on
hav ng several goals none of -before accept ng a better
wh ch can be ach eved w th cand date path, the front er nodes
certa nty are tested
-In real ty, part al observab l ty -> -front er s pr or ty queue based on
max m z ng expected ut l ty the r opt mal-path-cost
-before accept ng better cand date
path,
-the front er nodes are tested.

f le:///D:/Art f c al%20Intell gence.html 1/17


06.01.2019 Art f c al Intell gence.html

14. Propert es of Complete - Yes f step cost > 23. Greedy Search expand the node wh ch s closest
Un form-Cost ? eps lon (otherw se t can stuck n ? to goal n terms of heur st cs
nf n te loops) - t only takes the heur st cs not
T me - b[C*/e] # of nodes w th path the path cost
cost =< cost of opt mal solut on
Space - b[C*/e] # of nodes w th 24. Propert es of Complete - No(can stuck n loop)
path cost =< cost of opt mal Greedy search T me - O(bm) good heur st cs
solut on mprove t dramat cally
Opt mal - Mostly Yes for any step Space - O(bm) keep all nodes n
cost >= memory
Opt mal - NO
15. Depth-f rst L fo - last n , f rst out
Search ? Expand the deepest unexpanded 25. A* Search ? Evaluat on funct on f(n) = g(n) +
node
h(n)( g= cost to node, h=est mated
cost from node to goal, f=
16. Propert es of Complete - No ( h gh probab l ty t est mated total cost through node
Depth-f rst falls nto loops) to goal
search T me - O(bm) m=max mum depth Algor thm selects the m n mum
f m>>d (d= depth of least cost f(n) and expand t
solut on) -same as un form goal test
Space - O(bm) performed when node s selected
Opt mal - NO for expans on

17. Depth-l m ted depth-f srt search w th depth l m t l 26. Propert es of A* Complete - Yes, unless there are
search? Search nf n tely many nodes w t f(n) =<
f(G)
18. Iterat ve w th l m ted depth 0,1,2,3,...,d T me - Exponent al bd
deepen ng terate over d to f nd solut on Space - Keeps all nodes n
search IDS? -prefered when search space s memory
large and solut on depth s Opt mal - Yes
unknown (
- terat ve deepen n may seem A* expands all nodes w th f(n) <
wasteful because t generate early C*
nodes nult ple t mes but for deep A* expands some nodes w th f(n)
searches t s l ttle amount = C*
A* expands no nodes w th f(n) >
C*
19. Propert es of Complete - Yes )
IDS? T me - (d + 1)b0 + (d)b1 + (d-1)b2
+ ... +bd = O(bd) 27. Opt mal ty of A*
Space - O(bd)
Opt mal - Yes f step cost = 1 or
ncereas ng funct on of depth

20. Summary of
algor thms

28. Adm ss ble - A heur st c h(n) s adm ss ble f


heur st cs for every node n, h(n) =< h*(n)
where h*(n) s trye cost to reach
the goal state from n
21. Informed - Best-f rst search - Adm ss ble heur st c never
Search - Greedy Search overest mates the cost to reach
Strateg es? - A* search the goal
- Inadm ss ble(pess m st c)
22. Best-f rst dent cal to un form-cost search f heur st c break opt mal ty by
search nstead of g and order of pr or ty trapp ng good plans on the fr nge
queue - Adm ss ble(opt m st c) heur st c
h(n) = heur st c funct on = est mate slowdown bad plans but never
the cost from node to the goal outwe gh true cost.
- Also h(n) >= so h(G) = 0 for any
goal G

f le:///D:/Art f c al%20Intell gence.html 2/17


06.01.2019 Art f c al Intell gence.html

29. Dom nance of If h2(n) >= h1(n) for all n(both 34. Grad ent Assume whe have some cost-
Heur st cs needs to be adm ss ble) Descent ? funct on : C(x1,...,xn)(C(x,y) =
then h2 dom nates h1 and s beter sqrt((x-a)2+(y-b)2)
for search. 1. Compute the grad ent
- And h(n) = max(h1(n),h2(n)) h(n)
s also adm ss ble and dom nates
h1, h2

30. Cons stency of 2. Take a small step downh ll n


heur st c the d rect on of the grad ent

Error execut ng latex.


Generated f le:
c:\users\cod ng\appdata\local\tem
p\ank _temp\tmp.tex
Have you nstalled latex and
dv png? = determ ned number f t
31. How to come - It can be der ved from relaxed s b g b g steps f t s small small
up w th vers on's solut on of the problem steps
adm ss ble - For the 8 puzzle problem f t s 3. Check f C(x1,...,x ',...,xn)
heur st cs? relaxed and p les can go
< C(x1,...,x ,...,xn)
anywhere heur st cs g ves the
shortest solut on, f t relaxed w th 4. If true then accept move, f not
p les move any adjent square then reject.
heur st cs g ves shortest solut on 5. Repeat
the opt mal solut on cost of a
relaxed problem s no greater than 35. S mulated Idea: escape local max ma or
the opt mal solut on cost of the Anneal ng ? m n ma by allow ng some "bad"
real problem moves. But gradually decrease
the r s ze and frequency.
32. Local Search In many opt m zat on problem, the -One can prove: If T decreases
Algor thms ? path to the goal s rrelevant; the slowly enough, then s mulated
goal state tself s the solut on anneal ng search w ll f nd a global
F nd a solut on and mprove opt mum w th probab l ty
t(opt m zat on) approach ng 1 (however, th s may
- H ll cl mb ng( grad ent take VERY long)
ascent/descent) Temperature s def ned to
- S mulated anneal ng
- Genet c Algor thms

33. H ll - Cl mb ng
?
S mulated anneal ng search ng for
a max mum. The object ve here s
to get to the h ghest po nt;
however, t s not enough to use a
s mple h ll cl mb algor thm, as
there are many local max ma. By
cool ng the temperature slowly the
global max mum s found.

f le:///D:/Art f c al%20Intell gence.html 3/17


06.01.2019 Art f c al Intell gence.html

36. Genet c -A successor state s generated by 41. Chess b~35(branch ng factor)


Algor thms? comb n n two parent states m n max? m~100(max mum depth of the
- Start w th k randomly generated state space)
states (populat on) 35100 = 2,55 x 10154 not feas ble
- State s represented as a str ng we do not need to search for
over a f n te alphabet (oftehn a every node. Alfa beta prun ng cuts
str ng of 0s and 1s) max mum depth to half 100/2 =
-Evaluat on funct on (f tness 50
funct on). H gher values for better 3550 = 1,56 x 1077 wh ch s more
states. feas ble
- Produce the next generat o of
states by select on, crossover and
mutat on

42. α–β prun ng? α s hte best value (to MAX


node)found so far off the current
path f V s worse than( smaller
than) α, Max w ll avo d t => prune
that branch

37. Games vs "Unpred ctable" opponent =>


Search solut on s a strategy
Problems

38. Types of games determ


chance
n st c
perfect chess, blackgam
nformat checkers, mon,
on go monopoly
mperfec battlesh ps, poker, β s the best value( to MIN) found
t bl nd t c tac scrabble, so far off the current path f V s
nformat toe worse than β ( more than β ). MIN
on w ll avo d t => prune that branch

39. M n max ? It s perfect for determ n st c,


perfect- nformat on games( chess,
go..)
Idea: choose move to pos t on w th
h ghes m n max value = best
ach evable payoff aga nst best
play( t may not be su table aga nst
sub-opt mal player)

40. Propert es of Complete - Yes, f tree s f n te


m n max ? (chess has spesc f c rules for th s)
Opt mal - Yes, aga ns an opt mal
opponent. If sub-opt mal player
even though there s h gher
43. Propert es of α– Prun ng does not affect f nal
opt mal outcome m n max does
not go for the opt mal. β prun ng ? results
Good move oder ng mprove
T me complex. - O(bm) effect veness of prun ng
space complex ty - O(bm) (depth- Perfect order ng cuts the
f rst explorat on) max mum depth(m) to the half m/2

44. Expect m n max g ves good play for determ n st c


? and chance games
-It s same as M n max except
chance nodes needs to be
handled. Based on the probab l ty
g ven

f le:///D:/Art f c al%20Intell gence.html 4/17


06.01.2019 Art f c al Intell gence.html

45. How do we Standard approach: 49. Does False ⊨ Read as : False log c ally enta ls
make α–β • cutoff test: (where do we stop True ? True f all models that evaluate
pract cal? descend ng the tree) False to True also evaluate True
– depth l m t to True.
– better: terat ve deepen ng no model makes "false" true,
– cutoff only when no b g changes "False ⊨ True" s actually true! It's
are expected to occur next an nstance of vacuous
(qu escence search). mpl cat on: th nk of t as be ng true
• evaluat on funct on for the same reason the statement
– When the search s cut off, we https://math.stackexchange.com/q
evaluate the current state by uest ons/1721140/does-false-
est mat ng ts ut l ty. Th s est mate enta l-true-and-v ce-versa
s captured by the evaluat on
funct on. 50. Declarat ve Tell t what t needs to know
– Run α-β prun ng m n max w th
approach to Ask t what to do -> answers
these est mated values at the
buld an agent: should follow from the Knowledge
leaves nstead.
Base
Evaluat on funct ons:
• For chess, typ cally l near
we ghted sum of features Eval(s) 51. Why Do We • Problem-solv ng agents were
= w1 f1(s) + w2 f2(s) + … + wn Need Log c? very nflex ble: hard code every
fn(s) poss ble state.
• e.g., w1 = 9 w th f1(s) = (number • Search s almost always
of wh te queens) – (number of exponent al n the number of
black queens), etc. states.
• Problem solv ng agents cannot
46. Knowledge
nfer unobserved nformat on.
-a l st of facts that are known to
• We want an algor thm that
base the agent.
reasons n a way that resembles
- set of sentences n a
reason ng n humans.
formallanguage

52. log cs, Syntax, Log cs are formal languages for


47. What -Represent states, act ons.etc.
Semant cs represent ng the nformat on such
Knowledge- -Incorporate new percepts
that conclus on can be drawn
based agent -Update nternal representat on of
Syntax def nes th senteces n the
must be able to: the world
language
-Deduce h dden propert es of the
Semant cs def ne the "mean ng" of
worold and deduce appropr ate
sentences, .e. whether a
act ons
sentence s true n a world

48. Enta lment ? Means that one th ng follows from 53. Models n log c? In log c, models are formally
another
structured worlds w th respect to
KB ⊨𝛼
w ch truth can be evaluated
Knowledge base KB enta ls M s a model of sentence 𝛼 f
sentence 𝛼
sentence 𝛼 s true n M
f and only f a s true n all worlds
KB ⊨𝛼 f and only f M(KB) ⊆ M(𝛼)
where KB s true
( f there s any knowledge n KB
wh ch makes 𝛼 false then we say
KB ⊨𝛼 s false
f there s not enough knowledge
to make 𝛼 always true n KB we
say KB ⊨𝛼 s false)
Note f 𝛼 ⊨ 𝛽 then 𝛼 s a stronger
assert on than 𝛽
https://www.youtube.com/watch?
v=2M-K5OjVgYQ

f le:///D:/Art f c al%20Intell gence.html 5/17


06.01.2019 Art f c al Intell gence.html

54. Inference ? KB ├ α => sentence α(alfa) can 58. Propos t onal Syntax
be der ved from KB by procedure log c Propos t onal log c s the s mplest
Consequences of KB are log c - llust rates bas c deas
haystack; α s a needle
Enta lment = need n haystack;
nference = f nd ng t

55. Soundness? s sound f whenever KB ├ α, t s


also true that KB╞ α (no wrong
nferences, but maybe not all Semant cs
nferences) Each model spec f es true or false
In other words: for each propos t on symbol
A sound nference algor thm E.g. P1,2 P2,2 P3,1
der ves only enta led sentences false true false
(an unsound algor thm makes W th these symbols, 8 poss ble
th ngs up, .e., can d scover non- models, can be enumerated
ex st ng “needles”) automat cally.
When nference algor thm takes Rules for evaluat ng truth w th
alfa sentence from KB and KB respect to a model m:
enta ls alfa sentence t means ¬S s true ff S s false
nference algor thm s sound S1 ∧ S2 s true ff S1 s true and S2
s true
S1 ∨ S2 s true ff S1 s true or S2 s
56. Completeness? s complete f whenever KB╞ α, t
true
s also true that KB ├ α (all
S1 ⇒ S2 s true ff S1 s false ∨ S2
nferences can be made, but
maybe some wrong extra ones as s true
well) .e.,S1 ⇒ S2 s false ff S1 s true ∧
In other words: S2 s false
The nference algor thm s S1 ⇔ S2 s true ff S1⇒S2 s true ∧
complete f t can der ve any S2⇒S1 s true
sentence that s enta led from S mple recurs ve process
knowledge base. evaluates an arb trary sentence,
If algor thm can get every poss ble e.g., ¬P1,2 ∧ (P2,2 ∨ P3,1) = true ∧
sentece from knowledge base (true ∨ false) = true ∧ true = true
wh ch enta ls a then algor thm s
complete
59. Truth tables for
connect v t es?
57. How do we If Kb s true n real world, then any
know that sentence alfa der ved from KB by
knowledge sound nference procedure s also
base s true n true nthe real world. 60. Val d ty n log c A sentence s val d f t s true n all
real world? KB may not be true n the real models
world, t depends mostly on e.g., True, A ∨¬A, A ⇒ A, (A ∧ (A
learn ng, wh ch s fall ble. W th ⇒ B)) ⇒ B
good learn ng approach KB may Val d ty s connected to nference
be true enough for pract cal v a Deduct onTheorem:
appl cat ons. KB ╞ α f and only f (KB ⇒ α) s
val d

61. Sat sf ab l ty n A sentence s sat sf able f t s true


log c n some model
AVB
A => B
A sentence s unsat sf able f t s
true n no models
A∧¬A
Sat sf ab l ty s connected to
nference v a:
KB ╞ α f and only f (KB ∧¬α) s
unsat sf able (there s no model for
wh ch KB=true and s false)

f le:///D:/Art f c al%20Intell gence.html 6/17


06.01.2019 Art f c al Intell gence.html

62. Proof methods Two ma n k nds of proof methods: 67. Common


Model check ng m stakes n
§Truth table enumerat on (always Quant f er
exponent al n the number of
symbols n)
§Improved backtrack ng
§Local search (h ll-cl mb ng, m n-
confl cts, s mulated anneal ng)
–WalkSAT algor thm (p cks a
symbol randomly and chooses
randomly between “m n-confl cts” 68. us ng Fol
and “random accept/reject” the
fl pp ng of the symbol)
Appl cat on of nference rules
Leg t mate (sound) generat on of
new sentences from old ones
A proof s a sequence of nference
rule appl cat ons
–Search ng for proofs s an
alternat ve for enumerat ng
models 69. There s person
–Can be more eff c ent because who loves
the proof can gnore rrelevant everyone n the
propos t ons Inference rules world
requ re translat on of sentences Translate t to
nto a normal form FOL

63. CNF ? Conjunct ve Normal Form 70. Everyone n the


world s loved
64. Pros and cons (+)Propost onal log c s by at least one
of propos t onal declarat ve: p eces of syntax person
log c correspond to facts translate t to fol
(+)Propost onal log c allows
part al/d sjuncted/negated 71. Does No
nformat on(unl ke most data
structures and databases)
(+) Propos t onal log c s
compos t onal: A V B s der ved
from mean ng A and of B and
(+) Mean ng n propos t onal log c
s context- ndependent(unl ke
natural language where mean ng
depends on context are same?
(-) Propos t onal log c has very
l m ted express ve power (unl ke
72. Dual form of
natural language )
E.g. cannot say "p ts causes
breezes n adjacent squares" ?
except by wr t ng one sentence for
each square 73. Brothers are
s bl ngs
65. F rst order log c FOL assumes the world conta ns translate t to
FOL Objects(people, houses..), FOL
Relat ons(red, brother of, has
color), Funct ons(father of, best 74. "S bl ng" s
fr end..)
symmetr c
translate t to
66. Bas c elements Constants - K gnJohn, 2 FOL
of FOL Pred cates - Brother, >
Funct ons - Sqrt, LeftLegOf
75. One's mother s
Var ables - x, y, a,b,..
one's female
Connect ves - V, =>, <=>,..
parent
Equal ty - =
translate to FOL
Quant f ers - ∀ , ∃

f le:///D:/Art f c al%20Intell gence.html 7/17


06.01.2019 Art f c al Intell gence.html

76. A f rst cous n s 83. Why It s useful because log c fa ls due


a ch ld of a probab l st c to Laz ness(too many attr butes to
parent's s bl ng? reason ng? l st) and Theoret cal Ignorance( no
Translate t to complete knowledge of the
FOL doma n) and Pract cal Ignorance(
not enough, observat ons, tests,..)
77. PDDL? Plann ng Doma n Def n t on Porbab l st c assert ons
Language: common plann ng summar ze the effects of laz ness
language nstead of class cal and gnorance.
plann ng What probab l st c reason ng do
State: allowed and not allowed A cons stent framework for deal ng
states( At(fred,Gent),...) w th degrees of bel ef
Act ons: descr bed by a set of The pat ent w th ootache has a
act on schemas. cav ty w th probab l ty 0.8
Probab l t es relate propos tons t
one's own state of knowledge
78. 3 Cases to Mutex: mutual exclus on Probab l t es of propos t ons
occur mutex n change w th new ev dence
plann ng graph
84. How to choose
an act on?(what
s the elements
of dec s on
theory)

79. How mutex İf they negate each other: 85. What s In AI, propos t on s a set of
between l terals Have(Cake) and ¬Have(Cake) propos t on n poss ble worlds( events) where
occur? If any two act on comes to them AI? the propos t on holds
(y eld them) are mutex
¬Have(Cake) and ¬ Eaten(Cake)
86. P(a v b) = ?
are mutex because no act on and
Eat(cake) act ons are mutex

80. When the When 2 precond t ons are dent cal


plann ng graphs t levels of at the second one 87. Random --Propos t onal or Boolean random
levels off? var able types var ables
n propost on? Cav ty= true s wr ten as cav ty
81. how to show n Headache => S nus V Eyes ght V --D screte random var ables(f n te
log cal rule St ffNeck V Flu V Cancer or nf n te)
Pat ent w th Weather =
headache and {sunny,ra n,cloudy,snow}
many poss ble exhaust ve and mutually exclus ve
reasons: s nus --Cont n ous random vara bles(
problem, bounded or unbounded)
Eyes ght, Wa t ngT me= 381.2 also
St ffNeck, Flu, QueueT me <60
Cancer,...
88. What s Express bel ef g ven some
82. Why log c fa ls - There s too much work to l st all cond t onal ev dence
n probab l ty the attr butes probab l ty? P(cancer|smok ng) = 0.3 means
doma n - No complete theory or 0.3 g ven that smok ng s the only
knowledge ev dence for cancer
- Not all the necessary tests can when we know more we m ght get
be or have been run P(cancer|smok ng,-cough) = 0.5
rrelavant ev dence allows
s mpl f cat on
P(cancer|smok ng,l verpoolw ns) =
P(cancer|smok ng) = 0.3
Cond t onal probab l ty shows n
cond t onal probab l ty table (CPT)

f le:///D:/Art f c al%20Intell gence.html 8/17


06.01.2019 Art f c al Intell gence.html

89. P(a|b) = ? 93. How to f nd


cond t onal
probab l ty f
there s h dden
var able?

alfa s normal zat on constant


wh ch makes the total probab l ty
equal to 1.

94. How jo nt
d strub ton
90. repsentat on
change when
cond t onal
=? ndependence
ntroduced?
91. What s cha n
rule?

92. P(toothache)

g ven the jo nt
probab l ty table 95. Bayes' rule ? Product rule P(a ∧ b ) =
what are P(a|b)P(b)
P(toothache)=?
P(cav ty V P(toothache) = 0.108 + 0.016 +
toothacge) = ? 0.012 + 0.064 = 0.2
P(¬cav ty|tootha P(cav ty ∨ toothache) =
che) = ?
P(Cav ty|tootha
che) = ?
96. naïve Bayes
model?
P(cav ty ∨ toothache) = 0.108 +
0.012 + 0.016 + 0.064 + 0.072 +
0.008 = 0.28
P(¬cav ty|toothache) = ? 97. What s Markov Each node s cond t onally
Blanket n ndependent of all others g ven ts
Bayesnet: Markov blanket:
parents + ch ldren 9 ch ldren's
parents

P(¬cav ty|toothache) = 98. Compact


cond t onal
d strub t on?

P(Cav ty|toothache) = ?

99. S mple query


on the burglary
network?
P(B | j,m) = ?

f le:///D:/Art f c al%20Intell gence.html 9/17


06.01.2019 Art f c al Intell gence.html

100. What s S mple nference task of P(X|e). 109. How Inference Suppose we have query of P(B |
poster or P(X|e) s called poster or by var able j,m)
marg nal? marg nal. It s poster or el m nat on Inference by enumerat on does
d strubut on of subset of works?
var ables, n th s part cular case w th var able el mant on we turn
th s subset s only one var able X probab l t es to the factors and

101. S mple quer es compute poster or marg nal P(X |


n nference E=e) and factors are vectors l ke
task og BN

102. Conjunct ve P(X1,X2|E=e)=


quer es ? P(X1|E=e)P(X2|E=e) when we wr te the new form

103. Opt mal Dec s on networks nclude ut l ty and comput ng r ght to left starts
Dec s on ? nformat on. requ red for here are the steps
P(outcme | act on, ev dence)

104. Value of Wh ch ev dence to seek next?


Informat on ?

105. Sens t v ty To f nd wh ch probab l ty values


Analys s? are most cr t cal
we po ntw se product f1 and f7
106. what are the S mple quer es but f1 need normal zat on
nference Conjunct ve quer es afterwards we po ntw se 2 factors
tasks? Opt mal dec s ons
Value of nformat on
Sens t v ty analys s
Explanat on

107. Why Inference Because of t s repeated PS: P(U,x,y,z) = f(U,x,y,z)


by enumarat on calculat on of probab l t es probab l ty of var able equal to the
s n eff c ent? factor form

108. Ma n dea of El m nat ng the repated 110. What s bel ef Ingerence by local mesage
nference by calculat ons by carry ng out propagat on? pass ng among ne ghbor ng
var able summat ons r ght-to-left (bottom- nodes
el m nat on? up) stor ng ntermed ate results The message can loosley
for later use nterpreted as "I(node ) th nk that
you (node j) are that much l kely

111. Where can Networks w thout loops


bel ef
propagat on
can exact
nference?

f le:///D:/Art f c al%20Intell gence.html 10/17


06.01.2019 Art f c al Intell gence.html

112. Bel ef 113. Bel ef


propagat on n propagat on n
cha ns ? Trees?

114. what s the dea 1.draw N samples fom a sampl ng


of Inference by d strub tons S
stochast c 2. compute an approx mate
s mulat on poster or probab l ty ^P
3. show th s converge to the true
probab l ty P

f le:///D:/Art f c al%20Intell gence.html 11/17


06.01.2019 Art f c al Intell gence.html

115. Other methods -Sampl ng from an empty network 122. Yes


for nference by -Reject on sampl ng: reject P(a,b,c) = P(a)P(b)P(c|a,b)
stochast c samples d saggre ng w th
s mulat on? ev dence
-L kel hood we ght ng: use Does a and b
ev dence to we ght samples ndependent n
-Markov cha n Monte th s covergent so a and b are ndependent th s
Carlo(MCMC): sample from a graph s d fferent s tuat on from other
stochast c process whose graphs as we can see
stat onart d strub t on s true
poster or 123. No.

116. What s the Bayes an networks -drected


common graph cal models Does a and b
graph cal Markov random f elds -not g ven c
model and the r d rected graphs ev dence are
ma n ndependent n
propert es? th s covergent
graph?
117. Does S mple No.
cha n n BN S mple path through a graph( or a 124. When can we If along the path there s
conta ns loops? s mple cha n) s sequence of say path from A L near node w th ev dence
vert ces and edges where no to B n BN s D vergent node w th ev dence
vert ces(and hence no edges) are blocked Convergent node w thout
repated. In other words, a s mple ev dence (and w thout ev dence
cha n conta ns no loops. n descendants)

118. No. P(a,b,c) = P(a)P(c|a)P(b|c) 125.


Does a and b
ndependent? != P(a)P(b) = > a and b are not Does X and Y
n th s l near ndependent are blocked?
cha n?
It s not blocked s nce there s
119. Yes. noth ng can block on the path
It s not blocked and t s not d-
separated
Does a and b
ndependent
g ven c 126. What s d- If we have at least one not
ev dence n th s separat on? blocked path
l near cha n? two nodes are not d-separated
and we can not guarantee that 2
nodes are ndependent g ven the
120. No. ev dences
P(a,b,c) = P(a | c)P(b | c)P(c) If all the paths are blocked
then we can say 2 nodes ar d-
Does a and b separated g ven ev dence and we
ndependent n a and b are not ndependent can guarangtee that they are
th s d vergent ndependent g ven the ev dence
cha n?
127. d-seperat on
121. Yes between node
sets?

Does a and b
are a and b are ndependent g ven c
ndependent ev dence
g ven c
ev dence n th s
d vergent graph

f le:///D:/Art f c al%20Intell gence.html 12/17


06.01.2019 Art f c al Intell gence.html

128. N t s nval d because t s 131. Factor graphs


cycl c t c. It does not make sense. n graph cal
Only makes sene when P(A) = model
P(B) = P(C) = 1

Does th s graph
s a val d
bayes an
network?

129. What s
explan ng
away?

H = Happy
S = Sunny
R = Ra se (money)
When you see happy person and
you see the sunny weather t s
less l ke to person got a ra se
When you see heppy person and
you see and t s not sunny
weather and t s more l kely to
have a ra se
https://www.youtube.com/watch? 132. Sum-product -Exact nference n trees and
v=pyxyYWNo8Qw algor thm cha ns
Expla n ng away s a common -Qu te good approx mate
pattern of reason ng n wh ch the nference on other graph cal
conrmat on of one cause of an models
observed or bel eved event -Eff c ent, fast computat on
reduces the need to nvoke -Reus ng computat ons for
alternat ve causes. The oppos te d fferent nodes n the graph
of expla n ng away also can -Bel ef propagat on algor thm's
occur, n wh ch the conrmat on of general zat on
one cause ncreases bel ef n
another. We prov de a general
133. Bel ef -Sum-product
qual tat ve probab l st c analys s
Propagat on -Max-sum
of ntercausal reason ng
Algor thms? -Loopy-bel ef propagat on

130. In wh ch graphs
134. Loopy bel ef Sum-product and max-sum
the exact
propagat on vers on are there
nference v a
algor thm
message
vers ons?
pass ng s
poss ble
135. Dynam c Every kalman f lter model s DBN
Bayes an but few DBNs are KFs; real world
Networks vs requ res non-Gauss an poster ors
Kalman f lters ?

136. Idea beh nd Ensure that the populat on of


Part cle f lter ng samples(part cles) tracks the h gh
l kel hood reg ons of the state-
space

f le:///D:/Art f c al%20Intell gence.html 13/17


06.01.2019 Art f c al Intell gence.html

137. What are the F lter ng: P(Xt|e1:t) 143. Representat on Feed ach ne w th raw data
nference tasks bel ef state - nput to the dec s on (feature) automat cally d scover
n temporal process of a rat onal agent learn ng representat ons
probab l ty Pred ct on: P(Xt+k|e1:t) for k> 0 To do that deep learn ng:
models? evaluat on of poss ble act on representat on learn ng w th
sequences l ke f lter ng w thout mult ple layers.
the ev dence s mple but non-l near modules at
Smooth ng: P(Xk|e1:t) for 0 < k < each level
1
better est mate of past states, 144. Key deas of Layers are not des gned by
essent al for learn ng representat on(f programmers
Most l kely explanat on: argmax eature) learn ng Learn ng data from data w th
x1:tP(x1:t| e1:t) general method
speech recogn t on, decod ng
w th no sy channel 145. Is t true that Yes. It s true and t s because
s ngle layer The perceptron s s mply
138. V terb Ma n dea explanat on: perceptrons separat ng the nput nto 2
Algor thm? https://www.youtube.com/watch? can solve only categor es based on the
v=6JVqutwtzmo l nearly threshold funct on at the f nal
separable node
cases? w1I1 + w2I2 < t
139. V terb algor tm -It s var ant of max-product
propert es? product alog rthm w th back- f the LHS s < t, t doesn't f re,
track ng. otherw se t f res. That s, t s
-It max m zes over the ncom ng draw ng the l ne:
messages to each node w1I1 + w2I2 = t
-In pract ce, comput ng products And because the we ghts and
of small numbers may become thresholds can be anyth ng, th s
nstable, so often the problem s s just any l ne across the 2
reformulated as f nd ng the d mens onal nput space
max mum of the logar thm of the
jo nt probab l ty, wh ch results n 146. Can we use No because XOR problem s not
replac ng the products by s ngle layer l nearly separatable
summat ons, and y elds the “max- perceptron for
sum” algor thm. The back- XOR problem?
track ng g ves the same
sequence as w th the “max-
product” algor thm

140. What s the Bu ld ng ntell gent v sual


goal of v sual processors that can learn from a
ntell gence? l m ted set of examples and
general ze to unknown scenar os,
much l ke our v sual sytem does

141. What humans Ma nly feedforward process ng


do n f rst 100-
200
m l seconds of
v sual
percept on
147. Why mult ple A mult layer neural network can
layers n deep d stort the nput space to make
142. Learn ng vs. Class cal mach ne larn ng learn ng? the classes of data l nearly
feature process mostly depends on separable
eng neer ng feature extract on. F nd ng the
r ght features s d ffucult, requ res
148. What s wrong -It requ res labaled data
expert knowledge, tun ng...
Because of th s bottleneck n the w th back- almost all data s unlabaled, need
process, the dream of pattern propagat on? a labor ntens ve work to label
recogn t on was replace hand- data by hand
eng neered features w th -The learn ng t me does not sclae
representat on(feature) learn ng well
t s very slow n the neural
networks w th mult ple h dden
layer
-It can get stuck n poor local
opt ma

f le:///D:/Art f c al%20Intell gence.html 14/17


06.01.2019 Art f c al Intell gence.html

149. what s cha n 157. Notat ons to


rule n back- express agent's
propagat on? preferences

158. What s Behav or descr ble as


150. How to update w th grad ent descent Rat onal max m zat on of expected ut l ty
the we ghts n preferences?
the network?
159. ax oms of ut l ty constra nts on rat onal
delta k s found w th theory preferences
backpropagat on synonym
alfa s our learn ng rate
160. Ax oms of ut l ty
151. Overcom ng Keep the eff c eny and s mpl c ty theory:
the l m tat ons of us ng a grad ent method for
of back- adjust ng the we ghts, but use t
propagat on for model ng the structure of
sensory nput
-Adjust the we ghts to mox m ze
the probab l ty that a generat ve 161. Ex stence of
model would have procuced the ut l ty funct on
sensory nput
-Learn p( mage) not (label |
mage)

152. In whch 162. M cromorts ? one-m ll onth chance of death


Convolut onal neural networks
network back- useful for russ an roulette, pay ng
propagat on to reduce product r sks
works well
163. QALY ? Qual ty-adjusted l fe years
153. What s If we set the output to be the useful for med cel dec s ons
autoencoder? same as the nput, we can call nvolv ng substant al r sk
th s nural network an (pat ent w th a d sab l ty may be
autoencoder because t encodes w ll ng to accept a shorter l fe
a more dense representat on of expectancy to be restored to full
the nput data health e.g. years on d alys s vs.
shorter years at full health
154. Deno s ng Autoencoder can overf t and n
164. EMV(L) ? Expected monetary value of L
Autoencoder? order to force the h dden layer to
d scover more robust features
and prevent t from s mply 165. How dec s on
learn ng the dent ty, tra n the networks are
autoencoder to reconstruct the structured
nput from a corrupted vers on of
t.
So we ad no se to data(way of
regula ng and b ased data) and
tra n w th tat data
166. Expla n
Stochast c
155. Neural network Autoconders dom nance?
for Restr cted Boltzman mach nes
unsuperv sed
learn ng?

156. DEc s on- It makes dec s ons based on


theoret c what t bel eves and what t
agents wants. 167. What s the The agent's preferences have no
-comb nes bel ef and des res worst case n regular ty at all
under uncerta nty rat onal
- t has a cont n ous measure f dec s on
outcome qual ty
- the agent's des re s descr bed
by ut l ty funct on

f le:///D:/Art f c al%20Intell gence.html 15/17


06.01.2019 Art f c al Intell gence.html

168. Value of 175. How to be -Take correct f rst act on


Informat on for opt mal -Keep be ng opt mal
O l company accord ng to
bellman
equat on

176. Two ma n Value terat on


approaches for Pol cy terat on
comput ng
opt mal
pol c es?
169. Value of perfect
nformat on CPI
177. Problems w th It's slow - O(S2A) per terat on
bellman The "max" at each state rarely
equat on change
The pol cy often converge long
before the values

178. Pol cy terat on In t al ze pol cy (p )


algor thm Alternate the follow ng two steps
170. 2 parts of Probab l ty theory(what agent unt l convergence:
dec s on should bel eve) pol cy evaluat on
theory(what Ut l ty theory(what an agent pol cy mprovement
agent should wants)
do) 179. Compar son of They compute the same th ng(all
value terat on opt mal values)
171. a) cho ce s obv ous, nformat on and pol cy In value terat on:
worth l ttle terat on every terat on updates both
(cho ce obv ous because the values and pol cy
for g ven Expected ut l ty means are far we dont track the pol cy but
probab l ty from eachother, nformat on worth max ng n the equat on mpl c tyly
d strub t ons for l ttle beacuse var ance of recomputes t
ut l t es g ven d strub t ons are more or less the In pol cy terat on:
ev dences same #notsure about the we do several passes that update
determ ne f nformat on value) ut l t es w th f xed pol cy ( each
cho ce s b) cho ce s nonobv ous, pass s fast because we cons der
obv ous or not nformat on worth a lot only one act on, not all of them)
and s (cho ce s not obv ous because after the pol cy s evaluated, a
nformat on means are close to each other, new pol cy s choosen (slow l ke a
worth a lot or and nformat on worth a lot value terat on pass)
l ttle because var ance s d fferent) The new pol cy w ll be better(or
c) cho ce s nonobv ous, we converged)
nformat on worth l ttle Both are dynam c programs for
(cho ce nonobv ous because solv ng MDPs
means of probab l ty d strub t on
s close to each other and 180. Re nforcement Explorat on: you have to try
select ng s some how not easy, learn ng unknown act ons to get
probab l ty d strub ton of each mportant deas nformat on
ev dence has same var ance so n MDP Expl tat on: eventually, you have
nformat on s not worth a lot) to use what you know
Regret: even f you learn
172. what s opt mal one that max m zes expected ntell gently, you make m stakes
pol cy n MDP ut l ty f followed Sampl ng: because of chance,
you have to try th ngs repeatedly
D ffuculty: learn ng can b much
173. What expl c t reflex agent harder than solv ng a known
pol cy def nes MDP

174. How to avo d F n te hor zon: 181. Bas c dea of -Rece ve feedback n the form of
deal ng w th Term nate after T steps Re nforcement rewards
nf n te rewards D scount ng 0 < y < 1 Learn ng -Agent's ut l ty s def ned by the
smaller y means smaller hor zon reward funct on
absorb ng state: guarantee that -Must (learn to) act so as to
for every pol cy a term nal state max m ze expected rewards
w ll be reached -All learn ng s based on
observed samples of outcomes

f le:///D:/Art f c al%20Intell gence.html 16/17


06.01.2019 Art f c al Intell gence.html

182. Re nforcement It s offl ne re nforcement learn ng


188. Q-learn ng colled off-pol cy learn ng
learn ng as
mdp propert es Caveats:
you have to explore enough
you have to eventually make the
183. Model-based Learn approx mate model based learn ng rate small enough
learn ng on exper ences Bas cally, n the l m t, t doesn't
Solve for values as f the learned matter how you select act ons
model were correct
Step 1:Learn empr cal MDP
189. Temporal B g dea: learn fom every
model
Step 2: Solve the learned MDP D fference exper nce
learn ng

184. Why does Because eventually you learn the


190. Log cal
Model-based r ght model
Equ valance
learn ng work

185. Why does Because samples appear w th


model free r ght frequenc es
learn ng work

186. What s the Goal: Compute values for each


goal and dea state under pol cy
of D rect Idea: Average together observed
Evaluat on? sample values
Act accord ng to pol cy
Every t me you v s t a state, wr te
down what the sum of d scounted
rewards turned out to be
Average those samples

187. Pros and cons Pros:


of D rect Easy to understand
evaluat on Does not need any knowledge of
T, R
Eventally computes the correct
aerage values us ng just sample
trans t ons
Cons:
It wastes nformat on about state
connect ons
Each state must be learned
separately So t takes long t me
to learn

f le:///D:/Art f c al%20Intell gence.html 17/17

Potrebbero piacerti anche