Introduction to AI - concepts

Question 1

current trends in AI

Answer 1

exploit the strengths of the computer and don’t model human thought processes
focus on particular tasks and not on solving AI as a whole
strong scientific standards

Question 2

conference series

Answer 2

Biennial International Joint Conference on AI
Annual National Conference on AI
Biennial European Conference on AI

Question 3

The birth of AI

Answer 3

1956, Dartmouth Conference

Question 4

symbolic AI

Answer 4

model knowledge and planning in a way that is understandable to programmers

Question 5

subsymbolic AI

Answer 5

model intelligence similar to a neuron

Question 6

Why did the initial expectations of AI not work out?

Answer 6

lack of scalability
difficulty of knowledge representation
limitations on techniques and representations

Question 7

Forward checking

Answer 7

Search
Keeping track of remaining legal values for unassigned variables
Terminate search if any variable has no more legal values

Question 8

Local search for CSP

Answer 8

search
Work with complete states
Allow unsatisfied constraints
Min-conflict heuristic: select a conflicted variable and choose the value that violates the fewest constraints

Question 9

Zermelo

Answer 9

Game-playing
One player can force a win or both players can force a draw

Question 10

Retrograde analysis

Answer 10

Zermelo, game-playing
Generate all possible positions
Mark all positions where A would win…

Question 11

Horizon effect

Answer 11

Game-playing
The catastrophy can be delayed by a sequence of moves that don’t make any progress

Question 12

Forward pruning

Answer 12

Game-playing
Alpha-beta only prunes search trees when it is safe to do so

Question 13

Null-move-pruning

Answer 13

game-playing
add a “null-move” to the search (assume the current player does not make a move)
if the null-move results in a cutoff, assume that making a move will do the same

Question 14

Iterative deepening

Answer 14

game-playing
repeated fixed-depth search which works well for transposition tables

Question 15

move ordering - heuristics

Answer 15

game-playing

domain-dependent heuristics:
capture moves first
forward moves first

domain-independet heuristics:
killer heuristics (manage a list of moves that produced cutoffs at the current level of search)
history heuristics (maintain a table of all possible moves; if a move produces a cutoff, its value is increased)

Question 16

transposition tables - what does an entry store?

Answer 16

game-playing

state evaluation value
search depth of the stored value
hash key of the position
best move from the position (optional)

Question 17

What techniques does AlphaGo use?

Answer 17

game-playing

deep learning
reinforcement learning
monte-carlo tree search

Question 18

Inductive learning

Answer 18

Machine learning

Learn a function from examples

Ignores prior knowledge
Assumes examples are given

Question 19

Credit assignment problem

Answer 19

Machine learning games

Delayed reward

Question 20

Knightcap

Answer 20

Machine learning games

Learned to play expertly in chess

Improvements over TD gammon:
Temporal difference learning with deep searches
Played against partners on the internet (instead of self-play)

Question 21

Simulation search

Answer 21

Machine learning games

Estimate the expected value of each move by counting the number of wins
At each chance node select one of the options at random
Make all move choices quicky

Often works well even if the move selection is not that strong - fast algorithm

Question 22

UCT Search

Answer 22

Machine learning games

Best-known formulation of MCTS
Combines a UCB-based tree policy with random roll-outs

Exploitation vs Exploration
Choose move that has been visited most often (reliability) not necessarily the one with the highest value (high variance)

Question 23

AlphaGo

Answer 23

MCTS, deep learning, reinforcement learning from self-play

MCTS: roll-out policy, learned evaluation function instead of real game outcomes

Question 24

Policy networks

Answer 24

Machine learning games

How “good” is it to move to a position
Input position
Output position

Question 25

Value networks

Answer 25

Machine learning games

How desirable is it to be in this position
Input position
Output single value

Question 26

AlphaGo Zero

Answer 26

Learned only from self-play
Much less training data

Question 27

forward chaining

Answer 27

knowledge

derive new facts from known facts

elementary production principle:
for every rule and a set of facts and a substitution (which maps the body to the set of facts) one can derive one proof step
can be iterated until no further facts can be derived

Question 28

Resolution principle

Answer 28

knowledge

backward chaining
for disproving a statement, assume its opposite and show that it leads to a contradiction

Question 29

RDF

Answer 29

knowledge

resource description framework
allows for deductive reasoning (given facts and rules, we can derive new facts)

opposite: induction
deriving models from facts

Question 30

ontology

Answer 30

knowledge

explicit specification of a conceptualization
encode the knowledge about a domain
form a common vocabulary and describe the semantics of its terms
logical theory

Question 31

OWL

Answer 31

knowledge

Web ontology language
syntactic extensions of RDF

Question 32

Freebase

Answer 32

knowledge

2000s, collaborative editing
no fixed schema
acquired and shut down by Google

Question 33

Wikidata

Answer 33

knowledge

goal: centralize data from wikipedia
collaborative
imports other datasets
one of the largest public knowledge graphs

Question 34

DBPedia

Answer 34

knowledge

extraction frmo Wikipedia
using maps & heuristics

together with YAGO one of the most used knowledge graphs

Question 35

NELL

Answer 35

knowledge

never ending language learner

input: ontology, seed examples, text corpus
output: facts, text patterns
large degree of automation
occasional human feedback

Question 36

Linked open data

Answer 36

knowledge

many dataset are publicly available and connected to each other
using standards like RDF andd URIs for identification of entries

Question 37

Which fields have contributed to AI research?

Answer 37

history

philosophy
mathematics
psychology
economics
linguistics
neuroscience
control theory

Question 38

Why is NLP so hard? highly ambiguous at multiple levels…

Answer 38

lexical: same word, different meanings
syntactic: same sentence, different interpretations
semantic: the interpretation depends on its context, requires understanding of our world
discourse: the meaning of a sentence depends on the previous sentences

Question 39

Traditional NLP tasks

Answer 39

word segmentation (divide the input text into small semantic entities)
part-of-speech-tagging (assign each word its most probable role in a sentence)
syntactic analysis (find the most probable grammatical interpretation of the sentence)
semantic analysis (find the most probable meaning of a sentece and resolve references)

Question 40

Word2Vec

Answer 40

NLP

every word is represented as a lower-dimensional, non-sparse vector
train a deep neural network in a supervised way, using context of a word as additional input

2 variants:
continuous bag of words
skip-gram

Question 41

POS tagging is a process of tagging words in a sentece based on…

Answer 41

NLP

the definition of the word (thesaurus)
the context of the word (sequence learning task, Hidden Markov Models, Conditional Random Fields)

Question 42

N-gram model

Answer 42

NLP

language model, uses only n-1 words of prior context
unigram, bigram, trigram

Question 43

GPT

Answer 43

General Pre-trained Transformer

Question 44

Loebner competition

Answer 44

Philosophy

Modern day version of the turing test

Question 45

Mitsuki

Answer 45

Philosophy

Also known as Kuki
Successor of Eliza
Five-time winner of the Loebner competition

Question 46

Sussman anomaly

Answer 46

Planning

Subgoals are not independent