Exam August 2022 Flashcards
What are the advantages of using predicate logic as a knowledge representation language?
Predicate logic is structured to lead to a conclusion, e.g.:
The cat the cat meows
The cat is standing at the food bowl
->
The cat wants food
Predicate logic has the ability to succinctly summarize the inference process above which does
that it is better than natural language. Statement logic is not effective as it is not used for
inference processes.
Give some examples of why AI research ended up in crisis in the 70s (first AI winter).
During the 60s and 70s, symbol grounding hypotheses were used as the basis for everything
human intelligence and therefore also for AI. The majority of all research was based on this
hypothesis and they promised too much with too little results. Lighthouse voiced its criticism against
The AI research before the British Parliament and after that the UK strangled its resources
exploratory AI research. The United States followed soon after.
Give examples of criticism of Newell & Simon’s logic program. (Logic Theorist, General Problem Solver
etc.)
The criticism was based on its relevance and application to real problems. Newell&Simon was
proponents of symbols as the basis of human intelligence and created their programs from it
this hypothesis and while their programs could be classified as “cool” by the standards of the time
it was difficult to see what they actually brought to the world in terms of areas of application.
What was Weizenbaum’s purpose in creating ELIZA?
Weizenbaum’s real purpose with ELIZA was to make a caricature of a psychologist, he wanted
show that a computer could not replace a human when it came to therapy, unfortunately for
Weizenbaum worked Eliza better than he had hoped and the program became instead
raised to the skies as proof that an intelligent machine could actually be made. By today’s standards
however, ELIZA is very limited, so he was actually right in the end.
a) Print which nodes are explored if we use iterative depth first search where we explore 3
depth at a time, ie 0-3, 4-7, etc. It is usually said that the root node is at depth 0. The target is node j. (1p)
b) In which case is it an advantage to use iterative depth first search?
a) a, b, c, d, e, f, g, a, b, c, d, e, f, g, h, i, j
b) Firstly, it is guaranteed to find the most cost-effective way when it combines
the advantages of breadth first-search and depth first-search. In addition, it is good in a cyclical
graph since, unlike depth first search, it does not run the risk of ending up in an infinite loop.
What is the difference between an uninformed and an informed search algorithm?
Uninformed Search:
Lacks knowledge about the problem beyond what is given in the problem definition.
Informed Search:
Uses additional information about the problem to make more informed decisions.
A greedy algorithm can get stuck in a local minimum. What does that mean?
A greedy algorithm may make choices that seem optimal at each step but lead to a suboptimal overall solution because it fails to consider the global context.
If the target is f, what does our fringe look like after expanding node e and adding its children too
- depth first (1p)
- breadth first (1p)
A: Depth first Current Fringe
E3 H4, C2
H4 J5,K5,C2
J5 K5, C2
K5 C2
C2 F3, G3
F3 (final goal) G3 (Will not go here as the goal is
already found)
Breath first Current Fringe
E3 F3, G3
F3 (The goal is reached) G3, H4 (Will not go here as the goal is
already reached)
Consider a robotic lawnmower (an intelligent agent) tasked with mowing a soccer field.
The agent can take grass length and training days into account. Describe the robotic lawnmower as “a
goal‐based agent”
Performance measure: The grass is cut in the right place and at the right time.
Environment: Soccer field, lawnmower, software
Actuator: The blades, wheels, motor, program for setting times
Sensors: Map, grass, text and numbers, objects in the road
Explain the environment for a self-driving car agent.
A self-driving car moves on roads, of course. Environment is roads, maps, users etc.
This environment is inaccessible as there can always be things on the roads that cannot
eg cows or other cars are predicted. The car does not know the consequences of its action until it sees
next state. It is a stochastic environment. In addition, it is non-episodic if you exclude
from the fact that the self-driving car knows its way through GPS and therefore constantly acts to reach
the goal/final destination
The course literature claims “It is reasonable to claim that AI is everywhere in society”, and discusses
several areas (areas). Name three areas.
The course literature mentions, among other things:
Biomedical applications, for example neuronally controlled prostheses and others
medical use.
Intelligent agents, for example “slave” robots, robot vacuum cleaners, robot lawnmowers, etc.
Natural language processing, for example voice recognition software such as Siri, Alexa, etc.,
search engines and translation programs.
What is Natural Language Processing? What is it used for? What is “parsing”? What is
generation?
NLP is the processing and “understanding” of words and sentences by an AI. It is used for
example in voice recognition programs like Siri and Alexa, in search engines, i
translation program. Simply anything that requires the program to be able to perceive naturally
language.
Parsing is the understanding of language components by an AI (and a human), for example
syntactic structure, morphology and the grammar of a language. Without parsing no input that can
of course with the program.
Generation is the output, i.e. what the program returns. It must be something that is
relevant to the user, partly so that we can understand through our own human parsing but
also that it is relevant based on our input.
In simple terms describe the computational processing of a hidden node in a neural network.
The hidden node multiplies the input value and weight value of a single node. It then adds the
computed value to the value of the other nodes. There is also an activation function, usually
the sigmoid activation function that is calculated and then it sends the result to the output
node if the value is enough for activation.
What is the key architectural property of “Deep” Neural Networks that distinguishes them from
simple Neural Networks like the Single Layer Perceptron?
Deep NN are multiple layers of neurons that are connected and create a “mind” together.
Since there are more neurons they are less affected by malfunction, i.e the system can go on
even in one neuron malfunctions - “sturdier”, they are more dynamic, less affected by
irrelevant stimuli/disturbing signals, and of course have higher computational power. This also
means that they need advanced computer power to function, a reason why they weren’t
possible in early studies of ANN despite the theoretical advances in the 80s.
How is the computational processing in feedforward Deep Neural Networks similar to processing
within the human Visual Cortex?
Just like in the visual cortex they visual input can be processed in sequences. Different
neurons process different parts of (for example) an image, some process colour, some
process facial recognition etc, and it is a deepening process so simply put: basics first, going
deeper and deeper until a full understanding/processing of the entire input can be done.
