– Reading Assignment on Intelligent Systems
3 Jan ’18 Due:
17 Jan ’18
Submission should be typed with answers
in a different font color. E.g.,
assignment as a MSWord or PDF file to the ECE/SYS6410 web page on http://moodle.oakland.edu
your file with an identifiable name such as CheokKaC_HW1.docx.
your horizon on the disciplines and context of artificial intelligence (AI) and
various disciplines that follow: computational intelligence, machine
intelligence, intelligent controls, etc.
your thoughts on the AI topics are applied to control systems.
become fascinated at the potential of future AI and Robotics
For general knowledge, search the
internet (Google or other search engines) with keywords. While
Wikipedia is generally a good start.
But there are many other sites that may have better info.
Read up articles, magazines, or books
on the subject
rigorous technical/math detail
Being an OU student, you have the
privilege to access a huge collection of scientific & technical articles http://library.oakland.edu/articles/ejournal.htm. Take advantage of receiving free prints as
You are to enjoy discovering the world
of AI along with rapid advances of today’s technologies.
a clearer concept, description, methods, application, highlights, potential,
fascination, complexity and future of the subject matters.
may find that it is complicated to encapsulate the subject matter, i.e., there
is no one simple concept.
the questions below. They are generally
subjective and meant to stimulate thoughts.
Artificial intelligent systems (50%)
Start with https://en.wikipedia.org/wiki/Artificial_intelligence. Read the article and look up its reference
links as necessary. Then add one to two
sentences for each of categories regarding what it can do or what it is. Problem a. provided a partial sample.
Reasoning and problem solving are important problems of AI.
Earlier AI algorithms that researchers developed were imitating the
step-by-step reasoning that humans use to solve problems or make decisions.
These algorithms were intended to solve puzzles, games or economic problems.
Knowledge representation is the representation of
information of world such as objects, their properties, categories and relation
between them; situations and events; causes and effects; etc., in a form in
which a computer can utilize to solve complex tasks.
Knowledge representation aids Artificial Intelligence (AI)
to solve complex tasks like diagnosis of medical condition of patients, having
dialogues in natural language, knowledge based search
and services (like Yebol), etc.
AI planning involves realization of strategies that are to
be executed by intelligent agents, which help them to visualize the future and
make predictions on how their actions would change it. Unlike classical
planning approach, in AI planning the agents do not assume that it is not the
only actor and adopt based on the actions of other agents.
AI planning helps autonomous robots, autonomous ground
vehicles and unmanned aerial vehicles to plan their trajectories of motion and
the actions it needs to take to maximize the chances of achieving the desired
Learning is the process of acquiring new knowledge or
modifying the existing knowledge. Machine learning facilitates computers the
ability to learn and improve automatically without being explicitly programed.
Machine learning finds application in optical character and
speech recognition, game playing, autonomous vehicles, financial market
analysis, internet fraud detection, machine learning control systems, etc.
Natural language processing facilitates intelligent agents
the ability to understand human language. Speech recognition, language
translation, analysis of text, language parsing and understanding, question
Natural language processing find application in machine
translation, question answering, information retrieval and text mining.
Machine perception is a form of AI that provides the
computer the ability to interpret data from the sensors, in a way that is
similar to the human perception of the world around them using their senses.
Machine perception have applications in speech recognition,
image recognition and object recognition.
Motion planning and object manipulation are subfields of AI
and are of immense importance to robotics. They help a robot or other
intelligent agent to learn about its surroundings and decide on how to get from
one point to another.
Social intelligence of AI
networks is an important topic because it helps the AI agents to take better
decisions by predicting the actions of other humans by reading and
understanding their motives and emotional state.
In game theory, social
intelligence is vital as the AI agent must be able to detect the emotional
dynamics of the human interaction and take decisions accordingly to improve the
Artificial creativity or
computational creativity is a sub field of AI that deals to replicate human
level creativity using a computer.
Artificial creativity finds
uses in computer story generation, musical creativity, etc.
Artificial general intelligence is an
intelligence that helps a machine to perform intellectual tasks that a human
and brain simulation:
Cybernetics and brain
simulation is an approach of AI that creates mathematical models and perform
theoretical analysis of brain to understand the principles behind its
functioning, its structure and development, how information is passed etc.
intelligence is an approach of AI that is based on high level symbolic
representation of problems, logic and search. This approach assumes that many aspects of intelligence can be achieved by the
manipulation of symbols. The nature of symbolic AI is to accept
symbolic input information and create symbolic output information.
Symbolic AI approach can be
found in computer expert systems (an expert computer emulates the
decision-making ability of a human expert.) which use a network that connects
symbols and rules. The expert system processes these rules to take decisions
and determines the additional information it needs.
This method is easier to easier to explain, easier to
debug and better for abstract problems.
AI approaches intelligence without specific representations of knowledge. In this approach, information is parallelly
processed by mathematical calculations (example neural networks, Bayesian
learning, deep learning etc.).
approaches are more robust against noise, require less knowledge upfront,
easier to scale and is more useful for connecting to neuroscience
Statistical approaches or
evidence based approaches in AI involve using sophisticated mathematical tools
to solve specific sub problems. It uses optimization techniques to
automatically improve the performance of a piece of software, based on evidence
present in measurement data.
Statistical approaches can
handle uncertainties better than the symbolic and sub-symbolic approaches.
Some simple AI agents which
solve a specific problem can use any approach which works, i.e., symbolic,
sub-symbolic or statistical or any new approaches. Compilated agents like an
autonomous robot network, autonomous vehicles etc., will often have to take
approaches that integrate all these methods to be able to perform the complex
tasks that they would be given.
16. Search and optimization:
Search algorithms helps to
solve many problems in AI by searching through viable solutions. For real world
problems, simple search algorithms are rarely sufficient as they are either too
slow or never completes. The solution for real world problems is to use search
algorithms based on optimization techniques. Search algorithms retrieve
information stored within some data structure and the optimization techniques
allow the computer to select the best element, based on some criterion, from
set of available alternatives. This method can be thought of beginning with
some form of a guess (random or intelligent) and the continuously refining the
guess an incremental basis, until there is no further scope for refinement.
Optimization based search
technique can be found in path-planning of autonomous vehicles or robots, etc.
Logical AI uses set of statements that can be
true or false along with quantifiers and predicates, and express facts about
objects, their properties, and their relations with each other. Fuzzy logic is
an example of logical AI. Fuzzy logic based control systems can be found in
modern consumer products.
18. Probabilistic methods for
Probabilistic methods in AI employ probability theory and
economics to solve some problems in AI that have incomplete or uncertain
information. Some of the tools that are developed using these methods are
Bayesian networks, hidden Markov models, Kalman filter, particle filter, etc.
Bayesian networks find applications in image processing,
data fusion, risk analysis, document re classification, bioinformatics, etc.
Hidden Markov models find applications in computational finance, machine
translation, single molecule kinetic analysis, etc. Kalman filters find
applications in inertial guidance systems, nuclear medicine, weather
forecasting, etc. Particle filters find application in image processing,
molecular chemistry, computational physics etc.
and statistical learning methods:
classification and learning methods employ functions that use pattern
matching to determine a
closest matching category for the new data based on previous experience. Statistical learning theory deals with the problem of
finding a predictive function based on data. These learning methods have
become a powerful weapon to overcome uncertainty in AI scenarios and,
consequently, it has been widely implemented in many modern AI frameworks.
Applications include data
mining, e-mail spam filtering etc.
20. Neural networks:
Artificial Neural Network’s (ANN) are inspired by biological
neural networks and consists of interconnected group of neurons, where each
neuron is represented by a mathematical model. A trained ANN determines output
response to the input signal depending on the training values and the
mathematical functions of the neurons. Main categories of neural network are
feedforward neural networks and recurrent neural networks (allows feedback and
short-term memories of previous inputs).
ANN find applications in system identification, control
theory, pattern recognition, automated trading systems, quantum chemistry etc.
21. Deep feedforward neural networks:
Deep feedforward neural network consists of substantial
number of layers of neurons arranged in a feedforward way. They may contain many layers of non-linear hidden
units and a very large output layer. These networks are trained one
layer at a time. The recent growth in computational power have resulted in
development of more efficient methods of training of deep networks.
Deep feedforward neural
networks find applications in automatic speech recognition, bio informatics,
mobile advertising, image restoration, autonomous driving, etc.
22. Deep recurrent neural networks:
recurrent neural networks have many layers of interconnected units of
artificial neural networks and the connections between units form a cycle. They
constitute to deep learning ANN algorithms in which learning is based on data
representations, as opposed to task-specific.
recurrent neural applications are applied by Google, Microsoft and Baidu to
improve the performance of speech recognition.
There are several languages that are developed specially for
programming of artificial intelligence algorithms such as Python, Prolog, Lisp,
etc. Apart from these, there are also several other general-purpose programming
languages like C++ and simulation tools like MATLAB and SIMULINK to develop
artificial intelligence algorithms.
24. Control theory:
Intelligent Control is a branch of control theory that use
various artificial intelligence computational approaches like neural networks,
fuzzy, neuro-fuzzy control, Bayesian control etc.
Intelligent control finds applications in flight control
systems, autonomous driving systems, oil drilling processes etc.
To promote research and
development in AI, there have been established several competitions and prizes.
The principal areas where there are general machine intelligence,
conversational behavior, data-mining, robotic cars, robot soccer and games.
AI in healthcare uses algorithms to analyze relationships between prevention or treatment
techniques and patient outcomes by analysis of complex medical data.
Usage of AI appeared several times recently, for example,
Microsoft’s system to assist doctors to find the right treatment of cancer,
IBM’s Watson which not only won a game show but also successfully diagnosed a
woman suffering from leukemia, autonomous robot that performed surgery, etc.
AI facilitates the creation and evolution of self-driving
cars and trucks. As of 2016, reports suggest that there are more than 30
companies that are utilizing AI in the creation of driverless cars, to name
some companies, they are Tesla, Google and Apple. These vehicles incorporate
complex AI algorithms, along with complex control strategies to perform
functions such as autonomous
braking, autonomous lane changing, collision prevention, navigation, mapping,
trajectory planning, etc. based on the available data.
28. Finance and economics:
In finance and economic sector AI is used to prevent fraud protection
by detecting abnormal behavioral patterns of users. AI is also used in book
keeping, stock investment, online trading, estimation of supply demand curves,
reduction of information asymmetry.
29. Video games:
AI is also used in video games to generate adaptive and
human like intelligent behaviors from non-playing characters. It finds usage in
computer board games like chess, go, checkers, poker players, scrabble, etc.,
and many other types of games
30. Education in AI:
In the recent times many private bootcamps, free programs
and paid programs have been developed to facilitate education in AI. A recent
study concluded that there is a shortage of about 1.5 skilled AI professionals.
These bootcamps have developed some programs to meet that demand like The Data
Incubator and General Assembly.
31. Partnership on AI:
Companies like Amazon, Google, Facebook, Microsoft and IBM
have recently established a non-profit partnership whose aim is to conduct
research, share best practices, create educational material, consult the
relevant third parties, respond to
questions from the public and media, etc. Partnerships like these will help advance the
field of AI by formulating its
best practices and serving as a platform about artificial intelligence.
Turing’s “polite convention”:
Alan Turing developed a
test in 1950, called the Turing test, which is a test of a machine’s ability to
exhibit intelligent behavior equivalent to that of a human. Turing proposed
that a human evaluator would judge natural language conversations between a
human and a machine that is designed to generate human-like responses.
In the Turing test, a human
would evaluate natural language conversations between a human and a machine and
the machine is said to have passed the test if the human evaluator cannot
reliably distinguish between the responses of the computer and the responses of
In the field of AI, a
technology is no longer considered as AI and is removed from the list if it
reaches mainstream use. This phenomenon is called AI effect.
Task 2: Deep learning as an essence of Intelligent
Control System (50%)
a Google Scholar search on “self-driving car” and select the link called
“end-to-end learning for self-driving cars,” by M Borjaski, et al. It should take you to Cornell University
Library page, displaying the titled paper.
Download the pdf. Or here’s the
direct link https://arxiv.org/pdf/1604.07316.pdf.
the paper as best as you can, without the a priori understanding of the
technique they used. It’s well written
with clear explanation if the experimental process, data acquisition, deep
learning training of convolutional neural network, successful on-road
verification, and detail explanation of how CNN layers tries to learn.
following video is complementary to the above paper. Enjoy.
Published on Apr 6,
NVIDIA CEO Jen-Hsun Huang describes how
the open platform NVIDIA DRIVE uses AI to advance the driving experience, from
infotainment to autonomous vehicles, and how the data gathered from
sensor-filled cars will be used to create comprehensive HD maps, at the GPU
understanding the paper and video, explain in a simple way how this is similar
to how a new driver learns to drive a car.
In other words, compare the above experiments to teaching a kid how to
drive. A paragraph of several sentences.
general, the process of a human driving an automobile consists of three steps,
sensing, planning and acting, i.e., sense and perceive the environment (road
conditions, temperature, etc.), make decisions in response to those
perceptions, and control the actuators (throttle, brakes, steering, gear, etc.)
to achieve the desired motion. This philosophy also applies to autonomous
driving. The reference paper proposes an end to end AI strategy using neural
network and deep learning for autonomous driving, with emphasis on the method
of its training.
strategy proposed was to sense the environment using 3 cameras and feed it to a
deep CNN. The CNN would calculate the desired trajectory and propose a steering
angle command to achieve this trajectory. This proposed steering wheel command
would then be compared to the reference steering wheel angle generated by the
human driver. The error between these two signals would be used to train the
CNN and the weights are adjusted in a direction which reduces this error. The
method of training the CNN that is similar to how we teach a new driver to
drive the car. A student driver drives the car in the supervision of an
experienced driver who sits in the passenger seat. The student driver senses
the environment and applies the control commands accordingly, while
continuously taking inputs from the experienced driver and learning from the
student driver starts by driving simple maneuvers (for example lane following)
in simple road conditions (for example dry roads with clear lane markings,
etc.) and simple traffic conditions and incrementally learns by supervision on
how to handle complex maneuvers and uncertain scenarios. This is similar to the
approach in the paper, where the neural network was first fed with simple data,
for example the data where driver was just staying in the lane. Data consisting
of more complex scenarios, for example roads without lane markings, residential
roads, parked cars, unpaved roads, tunnels, foggy conditions, low friction
surfaces etc. were incrementally fed to the CNN and made it learn.