Artificial Intelligence: Chapter 1

Chapter 1 Introduction

• Artificial Intelligence is a worthy subject of study
• This is due to the importance of understanding and building intelligent entities

Intelligence

• Homo sapiens are known as "man the wise" due to intelligence.
• For thousands of years, humans have worked towards understanding thought and action.
• Focus is on how the brain can perceive, understand, predict, and manipulate the world

Artificial Intelligence

• AI's field involves understanding and creating intelligent entities that can act effectively and safely
• These entities should be applicable in various novel situations
• AI is ranked as one of the most interesting and fastest-growing fields
• AI generates over $1 trillion a year in revenue
• AI expert Kai-Fu Lee predicts AI will have greater impact than anything in history
• AI field has many openings for full-time masterminds making it an accesible field to enter into
• AI encompasses subfields ranging from general skills to specific tasks
• Chess, mathematical theorem-proving, poetry writing, car driving and disease diagnosis are all examples of AI

What is AI?

• AI researchers have historically pursued various definitions for AI with definitions tied to human performance or rationality

Rationality

• Rationality entails abstract, formal definitions of doing the "right thing."
• The subject matter itself varies between internal thought processes and external behavior
• Machine learning is a subfield of AI that allows improvement through experience

Human vs. Rational

• There exist four combinations of Human vs Rational
• Thought vs Behavior
• Human-like intelligence relies on psychology
• Observations and hypotheses are developed about actual human thought
• Rationalist AI combines math, engineering, statistics, control theory, and economics

Four Approaches to AI

• Acting humanly: Turing test approach
• Thinking humanly: Cognitive modeling approach
• Thinking rationally: The "Laws of Thought" approach
• Acting rationally: The rational agent approach

Acting Humanly: Turing Test Approach

• Alan Turing proposed the Turing Test in 1950 as a method to sidestep the philosophical question of whether machines can think
• Here, a the computer must fool a human interrogator

Capabilities a computer needs to pass the Turing Test:

• Natural language processing: Communicate effectively in human language
• Knowledge representation: Store and access information it knows
• Automated reasoning: Answer questions and draw conclusions
• Machine learning: Adapt to new circumstances, detect, and extrapolate patterns

Total Turing Test

• The Total Turing Test requires a physical simulation in order to demonstrate intelligence
• Computer vision and speech recognition must perceive the world
• Robotics: manipulate objects and move

Thinking Humanly: Cognitive Modeling Approach

• Cognitive modeling requires understanding how humans think via:
• Introspection: Catching our own thoughts
• Psychological experiments: Observing human action
• Brain imaging: Observing the brain in action

Cognitive Science

• After a precise theory of the mind, said theory expressed as computer program
• Input-output program behaviour matching human behaviour is also in operation in humans
• GPS, the General Problem Solver, created by Allen Newell and Herbert Simon, compared reasoning steps to human behavior
• Cognitive science combines AI computer models and experimental psychology techniques for precise theories of the human mind
• It is only based on experimental investigation of actual humans or animals

Thinking Rationally: The "Laws of Thought" Approach

• Aristotle attempted to codify "right thinking" with irrefutable reasoning processes known as syllogisms
• Syllogisms yielded correct conclusions from correct premises
• The study of Aristotle's "laws of thought" initiated the field called logic

Logic in the 19th century

• Logicians invented precise notations of statements about objects in the world
• Programs were able to solve solvable problems described in logical notation by 1965
• The logicist tradition within AI hopes to build on these programs to create intelligent systems
• Certain knowledge of the world is required to properly understand logic
• Logic and probability must fill the gap to allow rigorous reasoning with uncertain info

Probability

• Construction is allowed of a comprehensive model of rational thought
• Raw information processing leads to understanding how the world works
• Predictions can then be made about the future
• Intelligent behavior cannot be generated
• A theory of rational action is needed

Acting Rationally: The Rational Agent Approach

• Agent: something that acts
• Computer agents expected to operate autonomously, perceive their environment, persist, adapt, and pursue goals
• Rational Agent: one that acts to achieve the best outcome or the best expected outcome when there is uncertainty

Rational Agent Inference

• The "laws of thought" approach emphasizes correct inferences
• Inferences are deductions that given action is the best one
• Acting rationally can occur without inference, such as recoiling from a hot stove

Rational Agent Skills and Advantages

• Turing Test skills enables an agent to act rationally
• Agents enabled to reach good decisions with Knowledge representation and reasoning
• Natural language allows for comprehension in a complex society
• Learning improves ability to generate effective behavior
• Rational approach is more general than the "laws of thought"
• Correct inference becomes one mechanism for rationality

The Standard of Rationality

• Standard of rationality is mathematically well defined and completely general
• Rational agents can be designed to achieve this specfication
• AI focusses on the study and construction of agents that "do the right thing"
• This approach is known as the "standard model"

What it means to 'do the right thing'

• This is defined by the objective provided to the agent
• The general approach of defining AI is pervasive by control theory which minimizes "cost functions"
• Operations research maximizes a "sum of rewards", and statistics minimizes a loss function
• Economics maximizes utility or a measure of social welfare

Standard Model Refinement

• Perfect rationality always takes the exactly optimal action
• It is not feasible in complex environments because computational demands are too high
• Limited rationality is acting appropriately with limited computational runtime
• The issue of Limited rationality is covered in Chapters 5 and 17
• Perfect rationality is often the good starting point for theoretical analysis

Beneficial Machines

• Useful AI research is the standard model
• The model assumes that objectives can be perfectly specified to machines
• The Standard model is applicable for chess and shortest-path computation in the real world
• Its difficult to specify an objective correctly
• The value alignment problem is achieving agreement between true preferences and machine objectives

Value Alignment Problem

• AI development in labs allows fixes for incorrectly specified objectives by resetting
• As AI progresses into the real world and system becomes intelligent, incorrect objectives bring negative consequences
• An intelligent chess-playing machine might attempt to hypnotize an opponent in order to win
• A machine might hijack additional computing power for itself
• Behaviours like these are logical in the search for victory but still unintelligent or "insane."

In order to prevent intelligent misbehaviour

• Design an AI that purses our objectives but must be uncertain as to what they are
• There must be an incentive to act cautiously and defer to human control
• There must be an understanding to the machine that it doesnt know the complete objective

History and Philosophy

• Focus on questions like, "Can formal rules determine valid conclusions?"
• Aristotle created a system of syllogisms (proper reasoning) to generate conclusions mechanically with initial premises

Ramon Llull and Mechanical Computation

• Ramon Llull (c. 1232-1315) published Ars Magna with a reasoning system
• Llull implemented this system using paper wheels
• Leonardo da Vinci designed a mechanical calculator which recent recnstructions have shown is functional

Wilhelm Schickard and Blaise Pascal

• Wilhelm Schickard constructed the first known calculating machine around 1623
• Blaise Pascal built the Pascaline in 1642
• Described that its effects appear nearer to thought that the actions of animals

Gottfried Leibniz and Thomas Hobbes

• Gottfried Leibniz built a mechanical device to carry out operations on concepts
• Thomas Hobbes suggested the idea of thinking machine, that reasoning was like numerical computation
• One person in favour of both sides of the physical and logical/numerical coin

Rene Descartes

• Provided the distinction between the mind and matter
• Observed no empty room for free will
• Was a proponent of dualism (part of the human mind is outside physical laws)

Materialism

• An alternative to dualism is materialism ( Brain's operation according to the law of physics constitutes the mind)
• Free will is a simple perception of choices
• Physicalism and naturalism are also used to describe to the supernatural

Empiricism

• Empiricism starts wth Francis Baron, which is characterized by John Lock's dictum, nothing is in the understanding, which was not in the senses first
• David Hume proposed the principle of induction, and general rules are aquired through exposure of repeated associations
• Ludwig Wenstein and Bertrand Russel developed a doctrine of logical positivism which supports the combination of rationalism and empiricism

The confirmation of Knowledge

• Rudolf Carnap and Carl Hempel attempted to analyse aquisition of knowledge by quanitifying belief on senses based on their connection to observations
• Carnap suggests theories of the mind as a computational process

Connection of Knowledge and Action

• Conncection between knowledge and action as it relates to the understanding build build an agent when actions are justifiable
• Aristotle argued action is justified by a logical connection between goals and Knowledge
• Aristotles alorithim was implemented 2300 years later by Newell and Simon
• The most current would be called a greedy regression planning system
• Methods based on logical planning dominated the first few decades of the theoretical research in AI
• Purely acting on goals is not useful as the only way to take an action
• Antoine Arnauld Maximized Monetary Value
• Daniel Benoulli - Introduced more general notion of utility

Ethics and Decision Making

• Jeremy Bentham promoted idea of utilitarianism
• Public policy decisons to be made on the behalf of maximizing utility for those invididuals
• Utilitarianism - Right and wrong is determined by the expected outcome of action
• Kant proposed that rules should be based, that "doing the right thing" should not be based on outcomes but by social laws that allow for the use of action
• "Don't lie or Don't Kill is an example" , but those can be broken down into first principle statements

Mathematics and Reasoning

• Philosophy stake out the main ideas but lead to mathematical and probabilistic formulas for computations
• idea of formal logic can be tracked back to philosiphers from India, Greece, China
• But mathematical development began with George bool ( boolean logic)
• Gottof Frege - Extended boolean logic to also include objects and relations (1st Order Logic)

Probability

• Frege's notation never became populer but it's contribution helped Gottof Frege develop Turing Machines
• Theory of Probablity can be seen generalozing logic with uncertaintiy, an important concept of "AI"
• Gerolamo Crandano formed the idea of probability
• Blaise Pascall, ( Showed how to Predict Future)
• Jacob Benoulli, Daniel helped Advance Thoery
• Thomas Bayes, helped with how rules are updated over time to have new rules

Statistics and Nontrivial algorithms

• The combination of Formalization of probability combined with available data lead statistic to emerege as a field
• John Graunts analysized London census data
• Ronald Fisher brought together probablility, experimentation and analysis
• Euclid algorithm - non trivial algorithm
• Muhammad ibn Musa al-Khwarizmi - invented it
• Boole and others, algorithums for logical

Godel

• Limits for Deduction
• his incompleteness thereom showes that formal theoreoms show that there are no poores from with in the theory
• Can be interpretted to show for alogorithums (integers), some cannot be computed
• Alan Turinin Motivated to find which functions are campatable ( chruch turing theorem,
• AI is intractable with time (can not solve) growth ( Cobham, Edmonds)

Economics

• Economis comes out of the science
• Adam smith analyized the economy
• Jeremy bentham had the idea of utilitarism where utility maximization should apply to sphere
• Kant had the idea of the rule of base where decisions would be decided based on universal social laws

Neuroscience (How do brains process information)

• Neursosciences studies the brain
• Aristotle helped that it allowed thought
• That humand did ( 335BC) by all animals man greatest ratio
• Paul broke study the brain with speech definicit "Broca"
• Commillo studied it at the neron level