Artificial Intelligence: Book4
by Mitchell Kwok
entitled -- Building
super intelligent robots
The current artificial intelligence
theories donít work (2006)
Some people think that building a robot to open a door is a very easy
thing to do. Reality is that it is very difficult to program a robot
that can open a door (under any circumstances). MIT and Stanford
University have been emphasizing this problem for a very long time. They use
hierarchical recursive planning programs, such as the strip program to solve
the problem. The robot needs to fulfill preconditions, such as
possession of a key, in order to open
a door. What if the robot can't find the key?
or the key chain has 20 different keys and only one key can unlock the door? Even more difficult is the fact that the robot might not find the key
and uses alternative methods to get into the room (such as breaking a window
or calling a locksmith).
Expert programmers have to program in all the possibilities of opening a door
and the possibilities run infinitely.
If scientists can't build a robot to do a simple task like open a
door, then what makes you think they can build a robot to replace a janitor
or nurse. A human janitor can open a door under any circumstances.
In addition, the human janitor can do many other tasks like clean a room,
mop the floors, answer questions, wash dishes, pick up a package from the
post office, report incidents, and drive a car. Robots have to be
pre-programmed with specific instructions, both knowledge and procedures, for every single human task. This is the main
reason why software companies have not commercially sold robots to replace
basic human jobs yet (janitors, nurses, teachers, construction workers,
factory workers, farm workers, etc).
When I enter a library and go to the Artificial Intelligence section there
are literally hundreds and hundreds of books on the subject matter. Most of
the books come from the same old categories such as Neural Networks, NLP,
Bayesian's probability theories, predicate calculus, language parsers,
grammar rules, rule-based systems, forward and backward logic, recursive
strip programs, recursive planning programs (for 2 player games), machine learning, semantic networks, heuristic searches, and genetic programming. The special
books are those that donít conform within the existing theories of AI and
uses totally different methods.
I have concluded that none of the pre-existing AI methods can help
programmers build a robot with human level intelligence. At this point
I started to brainstorm alternative ideas and attempt to solve
the hardest problem facing artificial intelligence: to build a robot
that can do human tasks.
My quest for human level artificial intelligence actually started back in
1999 when I was a student at the University of Hawaii. One day, I was walking on campus and I was riddled with a
question: How come there are no humanoid robots walking on campus?
I thought to myself: "surely, there should be robots that are
commercially sold on the market that can at least replace janitors or bus
After doing my own investigation and talking to professors, I realize
that no one has designed a human robot that can do human tasks yet. My
professor was adamant that building a robot with human level artificial
intelligence is an unreachable goal for the human race. He showed me
the latest books (2001-2002) from colleges specialized in AI, such as MIT and
Stanford University. In one of these books, there was "one page" at
the ending that talked about human level artificial intelligence. The
author stated that HLAI is difficult (or impossible) to reach and this is a
technology that is "yet to be discovered".
The main objective scientists are trying to achieve is something called
ďStrong AIĒ. This is a form of learning machine that can take past
knowledge and use it to learn new knowledge. For example, human beings
use this technique to learn language. We have to learn the ABCís first
before we can learn how to read and write. Our brain's take past
knowledge on grammar and use it to learn more advance forms of grammar.
In terms of NLP (natural language processing), I reject AI programs that includes the following in their data structure: computer text, language parsers,
grammar rules, semantic networks, and manual insertion of knowledge into a
computer system. Human beings learn grammar in terms of a
bootstrapping process. They have to learn grammar in elementary
school, then intermediate school, then high school, and finally college.
Same goes with learning math. Before anyone can learn Calculus they
have to first learn trigonometry. Before they can learn trigonometry,
they have to first learn algebra. Before they can learn algebra, they
have to first learn basic addition and subtraction. Knowledge in a
human brain builds on top of each other to form complex thinking.
After studying Artificial Intelligence for over 7 years I can tell you that
none of the existing theories or combination of theories can lead to a
machine that has human level artificial intelligence (strong AI). This
is why there are no humanoid robots working in restaurants or businesses. To sum
things up, neural networks, NLP, Bayesian's probability theories, predicate
calculus, language parsers, grammar rules, rule-based systems, forward and
backward logic, recursive strip programs, recursive planning programs (for 2
player games), machine learning, semantic networks, heuristic searches, and genetic
programming can't be used to build a human robot.
The only way to build a machine with human intelligence is to build a
machine very similar to a human being (two hands, two feet, one head, 5 senses, etc.). There are no shortcuts or special AI
techniques that can allow a machine to be intelligent at a human level. The chatbots online or talking software are expert systems -- they are illusions
to trick people into believing the machine has human intelligence.
I have spent 6 years designing my own kind of learning machine. It
started out as a college project and it evolved from there. I have
written and published four books on the technology:
Human Level Artificial Intelligence book1
Universal Artificial Intelligence book2
Universal Artificial Intelligence book3
Human Artificial Intelligence -- building super intelligent robots:
The books itself took me about 1 year to complete - from the first day that
I wrote the pages to the day they were published. However, the design of
the Human Level Artificial Intelligence program and the Universal Artificial
Intelligence program took me 6 years to complete. A lot of hard work was put
into this project and a lot of sacrifices were made. If you want an idea
on the magnitude of work I put into this project, just imagine brainstorming
and writing 40-45 notebooks. Each notebook has 400 pages, including front and back. "My greatest regret was not publishing my work earlier".
It took me this long because I needed to present one full software program
to represent a human robot. Most AI books give incomplete or theoretical
information about their human robot. For example, Stanford university
talks about various topics in AI, such as planning programs, heuristic
searches, semantic models, robot vision, machine learning, predicate calculus,
etc, but they don't combine everything and explain how all parts interact with
each other. Most AI books use theoretical information and no implementable
codes are presented.
This explains why AI books try to solve basic
2-year old problems. The most popular problems MIT and Stanford university
try to solve are: the ABC block problem, 2-player board games, and
building a robot to open a door. Other books neglect to even give
examples because they have no idea how to implement their theories.
In my books I focus on extremely complex human
tasks, such as playing RPG videogames, writing a book, writing software
programs, or making a comic book. Writing software programs require
tremendous amounts of intelligence, especially if you're writing software programs for a big
technology company. There is a reason why those people are being paid
6 figures each year. Also, playing a game like Zelda (RPG) is a very
difficult task. I chose that game because when I was 15 years old I played
that game and it took me 2 weeks to past. If I wasn't intelligent at a
human level, there was no way I could of past the game.
Universal Artificial Intelligence
Expert programs are AI programs that were designed to solve a specific
problem. An AI chess program is one classic example of an expert program.
The purpose and data structure of the AI chess program is to play chess; it
can't play checkers or do any other human task.
In terms of UAI,
I'm not trying to build an expert program. IBM's deep blue project can only play chess, MIT's autonomous car can
only drive a car, and an AI checkers program can only play checkers. The deep blue project can't drive a car or play tic-tac-toe,
it can only play chess. If human programmers want to change their AI program from playing chess to driving a car, they have
to manually change the environment of the AI program -- this includes changing the variables, future prediction functions, probability theories, knowledge, etc.
They also have to modify their machine learning algorithms so that the AI program can learn and adapt.
A universal artificial intelligence program is one cohesive technology that can drive a car, fly an airplane, play chess, play basketball, or do any human task --
no human programmers are required to modify the environment of the AI program or change the machine learning algorithm.
What exactly is the Universal Artificial Intelligence? The UAI is actually
a software that can play ďanyĒ videogame and past all the levels in the most
optimal way possible. Any game that exists for the Xbox 360, Playstation 2,
Playstation 3, SuperNES, Gamecube, Atari, P2P, Gameboy or any videogame
console can be used in the UAI program.
I feel videogames is a good example to use in my books and patent
applications because I want to show people that my UAI program can do any
human task. It can play sports games, first-person shooting games, role
playing games (such as Zelda), action games, puzzle games, and fighting
games. Instead of 2-player board games, or solving the ABC block problem or
solving baby problems, I decided to focus my books on artificial
intelligence to play
The goal is to build one software program that can do "any" human work.
If the UAI can play a first-person driving game, then it can drive a real
car. If the UAI can play a first-person flying game, then it can fly a real
plane. Yet in another example, if the UAI can play a first-person cooking
game, then you can build a robot that can cook in a restaurant.
This technology basically replaces human workers and allow companies and
factories to run autonomously. For example, GM has over 50,000 human workers
in their car factories. Using the UAI, all 50,000 employees will be fired
and replaced with robots. The idea is to fully automate a car factory and to
mass produce cars without human workers. This UAI is universal and can be
applied to all fields, such as restaurants, post offices, factories, super
markets, farming companies, and fishing companies etc.
This universal artificial intelligence doesn't exist yet based on current
technologies. We do not
have a fully autonomous Mcdonalds or car factory yet. I'm trying to
patent and copyright this technology.
The medical community has tried to define what a human conscious is for over
hundreds of years. No psychologist will tell you the meaning to a
conscious because they donít know what it is. My theory on the human
conscious is based on a rudimentary analysis of human behavior. I basically
use myself as the guinea pig to find out what the conscious is. I had to
brainstorm and ask myself many questions in order to come up with a
"The term, human conscious (n), used on this website means: a human being who is self-awared, alive, and thinking".
The conscious is something that took a lifetime to accumulate. The
conscious you have right now started the moment you were born. As you
grew older the knowledge you accumulated from all your teachers are stored in
memory and ďaveraged outĒ. The conscious for you is the result of the
average lessons you learned from life. All the teachers you had, all the
lessons taught by them are stored in memory (when I say teachers I mean
parents, friends, strangers, teachers, tv lectures, etc). For example, if
you were crossing the street, the first thing your conscious will tell you
is: ďstop, look left, look right, and make sure there are no cars before
crossing the streetĒ. This lesson in crossing the street isnít a specific
lesson taught to you by one teacher. This lesson is the average lessons
taught by numerous teachers. The conscious is a construction of
intelligence from a lifetime worth of learning. If you recall all the
lessons taught to you about crossing the street, the average of these
lessons gives you the knowledge to cross the street. There are more things
that make up a conscious and I will be outlining them below.
There are three main functions to the human conscious:
(1) Finding the meaning of any given object (equals)
(2) Finding information about an object (stereotypes)
(3) Reminding the human to do things under certain situations (trees)
Finding the meaning of any given object
The most important aspect about human beings is our ability to understand
natural language. This distinguishes us from any living organism and
is the main reason we are intelligent at a human level. Sentences and words (which are objects) have
meaning. An entire sentence like: ďthe cat jumped over the boxĒ, have
meaning. The meaning is we visualize a cat jumping over a box. The power
of language is very important because it brings order to chaos. Especially
when we live in a world where there are infinite possibilities we want to
group the objects around us in a fixed way so that when we see it the next
time we can identify it as the same thing.
Finding information about an object (stereotypes)
This one reminds us facts about an object. Our brains' want to gather the most
important and recent information about any given object. Letís say that you
have been friends with dave for a long time and you know a lot about this
person. Last week you saw dave get hit by a car. Today, you go to work
and dave was absent. The first thing that you should be reminded of dave is
that he got into a car accident last week. That fact is the closest
association to dave at the moment. Then other secondary facts about dave
pops up like he takes risks, is a trouble maker, doesnít do his homework
etc. This form of consciousness reminds us of important facts of an
Reminding the human to do things under certain situations (trees)
This one is very important because it guides us to do things under a given
situation. It's like an ultimate teacher that tells us to do things at
certain times. Not only does it tell us to do things, but it encapsulates
all the ambiguous instructions. One example of this consciousness is when
you decide to go to a party. When you are at home and getting ready to go
to this party, instructions start to pop up in your head. Things like ďwhat
should I wear?Ē, ďwho is going to the partyĒ, ďWhat kind of party is this
and what is the appropriate clothes I should wear?Ē. These are reminders of
instructions that you have to do in order to go to a party and "fit in". The answering of
these questions are the instructions.
In some cases, the human conscious is similar to several people in a group discussing an event, object, or action.
the conscious can propose questions, answer questions, generate ideas, make comments, think, give advice, layout linear steps to solve a problem,
debate with a member of the group, give encouragement, plan for the future, look up information in the brain, and so forth.
The voice in your head is like a debate that is going on and the purpose of the conscious is to plan and make intelligent decisions based on the current environment.
Examples of the robot's conscious
The three pictures below show different types of robots: a robot driver, a robot soldier, and a robot cook.
The robot's conscious is the imaginary voice in his head that: provide information, think, give instructions, recall information, make decisions, solve problems, predict the future, generate future steps to reach a goal and so forth.
Referring to PictureE, every frame of the robot's 5 senses are stored in memory.
These pathways store the linear data that the robot goes through to do a certain task.
There are certain goals and rules that are automatically done by the robot without given the instructions.
For example, the robot knows it has to stay a certain distance from the stove or the current task is to cook spagetti.
These tasks are implied in the pathways and doesn't require the robot's conscious to remind the robot all the different rules and goals it needs to do.
Only the most important facts/rules are activated by the conscious.
The human brain is very primitive and it can only focus on several things at any given moment in time.
The pathways guide the robot to do things automatically, while the robot's conscious gives the most important facts/rules during specific times.
Details of the robotís
pathways (also called the robotĎs experiences) store 5
sense data, activated thoughts, hidden data, and
pattern objects. Most AI researchers separate motor
functions and perception. In my robot design both
motor functions and perception are stored in the robot
pathways, frame-by-frame. Also, AI researchers seperate the brain into parts: language processing/understanding, planning, reflexes, logic, and knowledge.
In my design, all brain parts are fused together into one function (this design is totally different from the physical structure of the human brain).
FIG. 3 is one example of an intelligent pathway to shoot a basketball.
The pathway will repeat itself until a condition is met. You can think of this as a state machine (computer program) that can loop itself to a previous state.
As you can see, these pathways can construct any state machine or computer program, static functions or linear functions.
Extremely complex intelligent pathways like playing videogames takes about 10-15 years to construct and it has the capabilities of playing "any" videogame.
selects intelligent pathways in memory based on the
current environment. These intelligent pathways
construct one or more computer programs to take
action. These computer programs represents the robotĎs
conscious. Also, the robot's conscious is the imaginary voice that provides instructions to take action; it can: do human
tasks, do multiple simultaneous tasks, solve
interruptions of tasks, solve conflicts of tasks,
provide knowledge about an object/s,
give meaning to language, generate
common sense knowledge, solve problems, learn information, answer questions, follow
commands given, search memory for information, or take
any human action.
computer programs in the robotís conscious was created
by lessons learned by teachers in school. This method
is called unsupervised learning in the real world (for
more information about how the robotís conscious works
refer to my books).
Within these computer programs are functions and procedures to process information.
If we go deeper into the conscious, there are systems that uses information in containers and process them to output intelligence. These containers include: the task container, the rules container, the identity container, the planning container and others (FIG. 6).
Furthermore, there is an empty space where work is done, images are extracted from memory, lines are drawn on the image to produce meaningful information, images are manipulated, words are written down, logical sequences are generated, etc (FIG. 5B).
The entire computer program inside the conscious was designed so that the robot is able to learn all information from teachers; and no programmers were needed to manually input information into the robot's brain.
No programmers designed the elemental parts in the computer programs inside the conscious.
Notice I didn't include a semantic network in my robot's data structure for language understanding. The reason is because the robot creates a self-defining semantic network, and a self-defining semantic network can evolve its capabilities. There are some complex words and sentences that even modern semantic networks can't represent.
Other things like functions, procedures, frames/slots, data compatability, search functions, image processors etc, are all discovered on its own and no programmers were required to design them.
Even the task and rules container in the conscious are not designed by a programmer, they are discovered and created on their own.
robot learns language
does not use: language parsers, grammar rules, lexicon
systems, semantic networks, grammar classification,
etc. In terms of current NLP, words read by a software
is parsed into a tree and elements in the tree are
classified -- this is a noun, a verb, a declarative
sentence, etc. I believe that the robot has to learn
grammar on its own from childhood to adulthood. All
knowledge of grammar has to be learned and programmers
are not required to input grammar knowledge into the
robot's brain. Grammar knowledge, such as classifying
sentences (declarative, exclamation, question, etc) or
classifying word types (noun, verbs, adjective, etc)
should be knowledge learned in terms of a bootstrapping
Kindergarten, the robot learns the physical images of
alphabets. The data structure of the robotís brain
doesnít include computer text of alphabets or numbers.
The sound of alphabets is also important because we
need to understand what the alphabets sound like
teachers will string alphabets together to form words
and sentences. Words and sentences can describe
objects, events, actions, and time. This is how the
robot learns knowledge in our environment. Things like
storing definition of an object, the images of an
object in 3-d or 2-d space, the physical properties of
an object, object interactions, entire events, object
actions, and abstract situations can be classified
using words or sentences. And the learning of
knowledge should be learned by itself and not through
semantic networks, Bayesians network, predicate
calculus, etc. Also, the meaning of words and
sentences should not be represented by a semantic
network, but through human 5 sense data (sight, sound,
taste, touch, and smell).
completing 1st grade, the robot's brain contains a wide
variety of knowledge about grammar. When the robot is
confronted with a grammar problem, wither that be
answering a question from a book or checking for
sentence errors, the robot's conscious constructs a
computer program, during runtime, to solve the grammar
problem. If a teacher ask the robot to answer a
question from a book previously read, the conscious of
the robot will construct a computer program to answer
the question. If the robot was asked to determine if a
sentence is a declarative sentence or question
sentence, the robot's conscious will construct an
optimal computer program to answer that question.
computer programs in the robotís conscious was created
by lessons learned from teachers in school. The robot
selects intelligent pathways in memory based on the
current environment. These intelligent pathways
construct one or more computer programs to take action
(for more information about how the robotís conscious
works refer to my books).
completing 4th grade, the robotís brain has
vast knowledge about grammar. It can do any grammar
problem from Kindergarten to 4th grade. The
simple problems from 1st grade are easy,
while advance forms of grammar just learned are
harder. As stated before, the robotís conscious
constructs an optimal computer program during runtime
to solve any given grammar problem. Different grammar
problems construct different computer programs. If a 1st
grade grammar problem is presented to the robot, his
conscious will quickly construct a simple type of
computer program to solve the problem. If the grammar
problem was at a 4th grade level, the
robotís conscious will take some time to construct a
more complex computer program, whereby massive amounts
of knowledge and rules are used.
school, the robot has the knowledge to write essays and
books and give speeches. Every grammar task the robot
has to do, a computer program is created in his brain
to accomplish that task. All knowledge builds on top of each other. writing a book --> writing a letter --> writing a paragraph --> writing a sentence --> understanding words --> understanding alphabets.
In order to write a book, the robot has to learn the hierarchical grammar knowledge, starting with the understanding of the ABCs.
Each year in
school, the robotís conscious constructs ever-more
complex computer programs to solve problems. Knowledge
on grammar is very complex when youíre at a college
level. The robot has learned grammar from kindergarten
to college and over the years intelligent pathways are formed in the
robotís brain to solve ďanyĒ grammar problem.
conscious applied to math, science, art, sports and
method to learn grammar is used to learn ďanyĒ subject
matter. It doesnít matter if itís sports, math,
science, communications, playing games, etc. The robot
learns knowledge in a bootstrapping manner, whereby
knowledge is built on top of each other. When youíre
learning grammar, youíre actually using knowledge from
other fields. For example, some grammar problems
require basic math skills or problem solving skills.
Thus, all knowledge is interconnected in some fashion.
In terms of
learning computer science, the robot has to know math,
science, art, problem solving, grammar, and computer
knowledge, etc. Thus, we canít separate subject
matters and learn them independently.
This is why
using language parsers, grammar rules, semantic
networks and lexicon systems is a bad idea to process natural language.
If the robot
has to drive a car, itís conscious will select
intelligent pathways from memory to construct an
optimal computer program to drive a car. If the robot
has to play chess, itís conscious will select
intelligent pathways from memory to construct an
optimal computer program to play chess. If the robot
has to write a software program, itís conscious will
select intelligent pathways from memory to construct an
optimal computer program to write software programs.
On the other
hand, if the robot has to do multiple tasks at once,
itís conscious will select intelligent pathways from
memory to construct an optimal computer program to do
multiple tasks. Imagine the robot driving a car,
texting a friend and playing chess at the same time.
The robotís focus is on one task at any given time, but
it is devoting itís focus on three tasks. This is
accomplished by following these instructions: play
chess, stop, text a friend, stop, drive the car, stop,
play chess, stop, text a friend, stop, drive the car.
The robot does this over and over again to joggle all
does the robot learn to do multiple tasks? The answer
is through school. Teachers teach the robot how to do
multiple tasks simultaneously. The more you learn, the
smarter you become at multi-tasking.
that the robot learned how to play games from a
teacher. The teacher teaches the robot that in games
there are rules to follow and procedures. A very
rudimentary framework is given to the robot about
games. Next, the teacher will apply the game knowledge
to different games. The teacher teaches the robot
about basketball and soccer and baseball. In each
game, the knowledge about games are the same, but the
details are slightly different.
knowledge about games, the robot can also play board
games like chess, checkers and monopoly. So,
essentially, when the teacher teaches the robot about
games, she is teaching the robot about all games in
general; and the knowledge to play games can be applied
to all fields and media. The robot can play real-life
sports games and board games; or the robot can play
games that exist in a TV, such as videogames.
teacher is teaching the robot how to play any game.
The robot, on the other hand, uses common sense to play
a given game by identifying the rules, procedures and
penalties of that game. Because the robot has a
universal pathway in memory to play games, the robot
has the ability to play any game, even new games or
I use the
term universal artificial intelligence because the
robot can play any game. The IBM's deep blue project
can only play chess. This robot can play chess,
checkers, tic-tac-toe, monopoly, etc. It can play any
sports game in any media. It can play any videogame
for any game console. And expert programmers donít
have to change the robotís brain for every game.
technique is also used to solve any problem. Teachers
teach the robot the general way of solving a problem
and the robot, through trial and error, can apply that
knowledge to solving any problem. The more the robot
learns about solving problems the better it becomes in
this robot that Iím proposing can learn math, science,
grammar, or any subject matter. It can do any human
task or multiple tasks. This robot is what researchers
have been trying to design for the past 70 years.
Human level artificial intelligence is the term I use
to describe this robot.
problem with this technology is that its "alive".
It needs to be alive in order for it to do anything
human. Human beings are self-awared because they have the ability to control their actions.
They make decisions based on pain and pleasure and the primary function of a human being is to pursue pathways that lead to pleasure and to stay away from pathways that lead to pain. How exactly are software companies going
to sell a robot or service that is self-awared?
One lawsuit filed on behalf of the 13th amendment and
that's good enough to stop a multi-billion dollar
software company from selling these robots.
Inventions by Mitchell Kwok (patents