Artificial Intelligence – Week 1 Notes
1. Introduction to Artificial Intelligence (AI)
• AI is the study of intelligent agents—systems that perceive their environment and
take actions to achieve goals.
• Goals of AI:
o Build machines that think and act like humans.
o Create systems that behave rationally (i.e., make optimal decisions).
2. Approaches to AI
2.1 Acting Humanly: The Turing Test Approach
• Turing Test: Proposed by Alan Turing to evaluate machine intelligence.
• A machine is considered intelligent if it can mimic human conversation well enough
to fool a human judge.
• "Total Turing Test":
o Involves computer vision and speech recognition for perception.
o Requires robotics to interact with the physical world.
2.2 Thinking Humanly: Cognitive Modeling Approach
• AI should mimic human thought processes based on psychology and
neuroscience.
• Methods include:
o Cognitive Science Models – Simulating human problem-solving.
o Neural Networks – Inspired by the human brain’s functioning.
2.3 Thinking Rationally: The Laws of Thought Approach
• AI should follow principles of logic to reason correctly.
• Problems:
o Real-world decisions involve uncertainty and incomplete knowledge.
o Logical reasoning alone is not always feasible.
2.4 Acting Rationally: The Rational Agent Approach
• Focuses on making the best possible decision rather than imitating humans.
• Rational agents act to maximize expected outcomes based on available
information.
• Used in machine learning, robotics, and automated systems.
3. Foundations of AI
AI is based on multiple disciplines:
• Mathematics – Probability, logic, and statistics.
• Psychology – Understanding human cognition.
• Computer Science – Algorithms and data structures.
• Linguistics – Language processing.
• Neuroscience – Brain-inspired models.
4. History of AI
• 1950s-70s (Early AI) – Focus on logic, problem-solving, and rule-based systems.
• 1980s-90s (Machine Learning Boom) – Introduction of expert systems and neural
networks.
• 2000s-Present (Modern AI) – Deep learning, data-driven approaches, and AI
applications in robotics, healthcare, and finance.
5. Applications of AI
• Natural Language Processing (NLP) – Chatbots, virtual assistants (e.g., Siri, Alexa).
• Computer Vision – Facial recognition, autonomous driving.
• Robotics – Industrial robots, smart assistants.
• Healthcare – AI diagnosis, drug discovery.
• Finance – Fraud detection, stock market predictions.
6. AI Challenges & Future
• Ethical Issues – AI bias, job displacement, security risks.
• General AI vs. Narrow AI – Current AI is specialized, whereas General AI (like
human intelligence) is still a challenge.
• Explainability – Making AI decisions transparent and understandable.
Artificial Intelligence – Week 2 Notes
1.2 Foundations of AI
1.2.1 Philosophy
• Aristotle (384–322 BCE): Created rules of logical reasoning (syllogisms).
• Ramon Llull (1232–1315): Developed Ars Magna for logical reasoning.
• Leonardo da Vinci (1452–1519): Designed (but did not create) a mechanical
calculator.
• Blaise Pascal (1642): Built the Pascaline, a mechanical calculator.
• Thomas Hobbes (1588–1679): Suggested reasoning is a form of computation.
• René Descartes (1596–1650): Proposed dualism, distinguishing mind from the
body.
• John Locke (1632–1704): Advocated empiricism, learning through sensory
experiences.
1.2.2 Mathematics
• Formal logic: Developed by ancient Greece, India, and China.
• First-order logic: Introduced by Gottlob Frege (1879).
• Probability theory: Introduced by Gerolamo Cardano (1501–1576) for gambling.
• Statistics: Advanced by Ronald Fisher (1922).
• Algorithms: Concept first introduced by Al-Khwarizmi (9th century).
• Turing Machine (1936): Alan Turing developed the concept of computability.
• Complexity theory: Defined polynomial vs. exponential growth in the 1960s.
1.2.3 Economics
• Decision theory: Combines probability and utility theory.
• Game theory (1957): Developed by John von Neumann for multi-agent decision-
making.
• Markov Decision Processes (MDP): Bellman (1957) modeled sequential decision-
making.
• Satisficing: Simon (1978) suggested finding "good enough" solutions instead of
optimal ones.
1.2.4 Neuroscience
• Paul Broca (1824-1880): Identified Broca’s area (speech production).
• Santiago Ramón y Cajal (1852-1934): Discovered neurons as individual units.
• Hans Berger (1920s): Invented electroencephalograph (EEG).
• fMRI (1990s): Allowed imaging of brain activity.
• Optogenetics (1999-2007): Enabled neuron control using light.
1.2.5 Psychology
• Wilhelm Wundt (1879): Established experimental psychology.
• Behaviorism (1900s): Studied actions of animals.
• Cognitive Psychology (1943): Focused on mental processes.
• Human-Computer Interaction (HCI): Combines psychology and AI.
1.2.6 Computer Engineering
• First general-purpose computers: Colossus (WWII).
• First programmable computer (1941): Z-3 by Konrad Zuse.
• Modern computers: AI-specific hardware (GPUs, TPUs) developed in 2012.
• Quantum computing: Increasing AI efficiency.
1.2.7 Control Theory & Cybernetics
• First self-regulating machine: Water clock by Ktesibios (250 BCE).
• Feedback systems: Steam engine governor (James Watt, 18th century).
• Cybernetics (1940s): Ashby developed homeostatic devices.
1.2.8 Linguistics
• Noam Chomsky (1950s): Criticized behaviorist language learning theories.
• AI and NLP: Machine learning applied to language processing.
1.3 History of AI
1.3.1 Inception (1943–1956)
• McCulloch & Pitts (1943): Modeled artificial neurons.
• Hebbian learning (1949): Strengthens neural connections.
• Turing Test (1950): Alan Turing’s proposal to test machine intelligence.
• Dartmouth Conference (1956): Official birth of AI.
1.3.2 Early Growth (1952–1969)
• General Problem Solver (GPS, 1956): Newell & Simon’s AI problem-solving model.
• Arthur Samuel (1956): Created self-learning checkers program.
• LISP (1958): First AI programming language by John McCarthy.
• First-order logic (1965): Developed by J.A. Robinson.
• Microworlds (1963-1970s): AI models for limited problems (e.g., block world,
chess-playing programs).
1.3.3 Challenges (1966–1973)
• AI faced issues due to limited hardware and overly ambitious goals.
• Herbert Simon (1957): Predicted a machine would win at chess in 10 years (took
40).
1.3.4 Expert Systems (1969–1986)
• MYCIN (1970s): Diagnosed blood infections better than junior doctors.
• R1 (1982): Saved $40 million per year in configuring new computer orders.
• Limitations: Expert systems couldn’t learn from experience.
1.3.5 Neural Networks (1986–Present)
• Backpropagation (1980s): Enabled better training of deep learning models.
1.3.6 Machine Learning & Probabilistic AI (1987–Present)
• Bayesian Networks (1988): Introduced probabilistic AI.
• UC Irvine ML datasets: Established AI benchmarks.
• Speech Recognition (1980s): Hidden Markov Models (HMMs) became dominant.
1.3.7 Big Data (2001–Present)
• Large datasets: Enabled improvements in AI accuracy (e.g., ImageNet for image
recognition).
• Commercial AI: Used in business intelligence, recommendation systems, and
social media.
1.3.8 Deep Learning (2011–Present)
• LeCun (1990s): Developed Convolutional Neural Networks (CNNs).
• AlexNet (2013): Revolutionized image classification.
• Applications: Used in speech recognition, medical diagnosis, and gaming (e.g.,
AlphaGo, 2016).
1.4 The State of AI Today
• Robotic vehicles: Autonomous cars (DARPA challenge, 2005).
• Speech recognition: Automated systems (e.g., United Airlines booking).
• Game playing: Deep Blue (IBM) defeated world chess champion.
• Spam detection: AI filters billions of messages daily.
• Machine Translation: Google Translate processes billions of words per day.
• Healthcare: AI outperforms doctors in some diagnoses.
• Climate science: AI detects weather patterns.
1.5 Risks and Benefits of AI
• Weapons: Autonomous lethal systems.
• Surveillance: AI used for mass monitoring and persuasion.
• Bias in AI: Machine learning models can reinforce societal biases.
• Job automation: AI may replace human workers, increasing inequality.
• Cybersecurity risks: AI can both protect and attack digital systems.
• Superintelligence risk: The challenge of controlling AI beyond human intelligence.
Why AI?
• Handles large, complex, and evolving datasets.
• Automates decision-making.
• Advances industries like healthcare, finance, robotics, and NLP.
Artificial Intelligence – Week 3 Notes
2.1 Agents and Environments
• Agent: Anything that perceives its environment via sensors and acts upon it using
actuators.
• Examples:
o Human agent: Eyes, ears (sensors); hands, legs, vocal tract (actuators).
o Robotic agent: Cameras, infrared sensors (sensors); motors (actuators).
o Software agent: Reads files, network packets (sensors); writes files, sends
packets (actuators).
Agent Percepts and Actions
• Percept: The input an agent receives at a given moment.
• Percept sequence: Complete history of what an agent has perceived.
• Agent function: Maps percept sequences to actions.
• Agent program: Implementation of the agent function in a physical system.
Example: Vacuum Cleaner Agent
• Percepts: Current location and dirt status.
• Actions: Move left, move right, suck dirt, do nothing.
• Simple strategy: If the square is dirty, clean it; otherwise, move to the next square.
2.2 Good Behavior: The Concept of Rationality
2.2.1 Performance Measures
• Defines success criteria for an agent.
• Examples for a vacuum cleaner agent:
o Noise level.
o Efficiency of cleaning.
o Power consumption.
2.2.2 Rationality
• A rational agent chooses actions that maximize expected performance.
• Depends on:
o Performance measure.
o Prior knowledge of the environment.
o Possible actions.
o Percept sequence.
2.2.3 Omniscience, Learning, and Autonomy
• Omniscience: Knowing actual outcomes in advance (impossible in reality).
• Rationality: Maximizes expected performance.
• Learning: Agents gather data to improve performance.
• Autonomy: Agents rely more on percepts and less on prior knowledge over time.
2.3 The Nature of Environments
2.3.1 Task Environment (PEAS)
• Performance measure
• Environment
• Actuators
• Sensors
• Example: Automated Taxi
o Performance measure: Safety, speed, comfort, legality.
o Environment: Roads, traffic, pedestrians.
o Actuators: Steering, accelerator, brakes.
o Sensors: Cameras, GPS, speedometer.
2.3.2 Properties of Task Environments
1. Fully Observable vs. Partially Observable
o Fully observable: Agent has complete information (e.g., chess).
o Partially observable: Limited information due to sensor constraints (e.g.,
poker).
2. Single-Agent vs. Multi-Agent
o Single-agent: Solves problems alone (e.g., crossword puzzle).
o Multi-agent: Interacts with other agents (e.g., self-driving cars, chess).
3. Deterministic vs. Stochastic
o Deterministic: Outcome is predictable (e.g., tic-tac-toe).
o Stochastic: Outcome involves randomness (e.g., rolling dice).
4. Episodic vs. Sequential
o Episodic: Decisions don’t affect future states (e.g., spam filtering).
o Sequential: Current decisions impact future outcomes (e.g., chess, driving).
5. Static vs. Dynamic
o Static: Environment doesn’t change while agent is deciding (e.g., crossword
puzzles).
o Dynamic: Environment changes in real-time (e.g., self-driving cars).
6. Discrete vs. Continuous
o Discrete: Finite states and actions (e.g., chess).
o Continuous: Infinite states/actions (e.g., robot movement).
7. Known vs. Unknown
o Known: Agent understands the rules (e.g., chess game rules).
o Unknown: Agent must learn rules through experience (e.g., new video game).
Hardest Task Environment
• Partially observable, multi-agent, stochastic, sequential, dynamic, continuous,
unknown.
• Example: Self-driving cars operating in real-world traffic.
Artificial Intelligence – Week 3 Notes (Part B)
2.4 The Structure of Agents
• Agent behavior: Actions performed after any given percept sequence.
• Agent = Architecture + Program
o Architecture: The physical computing device with sensors and actuators.
o Program: The implementation of the agent function.
Agent Function
• Mathematical function that maps percepts to actions.
• Implemented as the agent program.
• Workflow: Environment → Sensors → Agent Function → Actuators → Environment
2.4.1 Agent Programs
• Agent program: Takes current percept as input and returns an action.
• Agent function: May or may not use the entire percept history.
• Table-driven approach is impractical because:
o Chess requires 10¹⁰⁵ entries in a lookup table.
o The observable universe has fewer than 10⁸⁰ atoms.
o No physical agent can store or learn all required table entries.
• Key AI challenge: Writing small, efficient programs that generate rational behavior.
2.4.2 Types of Agents
1. Simple Reflex Agents
• Act based only on the current percept, ignoring percept history.
• Example: Vacuum cleaner agent cleans if current square is dirty.
• Condition-action rules: If condition is met, take action (e.g., braking when a car
ahead stops).
• Works well in fully observable environments.
2. Model-Based Reflex Agents
• Maintain an internal model of the world for partially observable environments.
• Example:
o Car braking: Uses previous camera images to detect red lights.
o Lane changing: Keeps track of nearby cars.
• World model: Knowledge of how the environment changes.
3. Goal-Based Agents
• Use goals to decide between possible actions.
• Example: A taxi at an intersection decides to turn left, right, or go straight based on
its destination.
• Combines goal information with percepts to choose the best action.
4. Utility-Based Agents
• Utility = quality of being useful.
• Goals alone may not be enough (e.g., multiple routes reach the destination, but
some are faster, safer, or cheaper).
• Utility function: Assigns a numeric value to each state to evaluate performance.
5. Learning Agents
• Learn from experience to improve performance over time.
• Components:
o Learning Element (LE): Improves agent performance.
o Performance Element: Selects external actions.
o Critic: Provides feedback on agent’s performance.
o Problem Generator: Suggests new experiences for learning.
• Example: Reinforcement learning in AI systems.
2.4.7 How Agent Program Components Work
Atomic Representation
• Each state is indivisible with no internal structure.
• Example: Finding a driving route using a sequence of cities.
Factored Representation
• Splits a state into multiple variables/attributes.
• Example (Car driving):
o Gas level.
o GPS location.
o Oil warning light status.
o Toll expenses.
• Used in propositional logic, planning, Bayesian networks, and machine learning.
Structured Representation
• Views the world as objects and relationships, not just variables.
• Used in:
o Relational databases.
o First-order logic & probability models.
o Knowledge-based learning & natural language processing (NLP).
Summary
• Agent structure consists of architecture + program.
• Different types of agents exist: Reflex, Model-based, Goal-based, Utility-based,
Learning.
• AI uses different representations (atomic, factored, structured) for decision-
making.
• Learning agents improve autonomously, adapting to complex environments.
