Introduction to AI

• AI encompasses the study of concepts enabling computer intelligence.
• AI is a field within computer science focused on creating systems with human-like intelligence.
• AI may be defined as a branch of computer science that allows the creation of intelligent machines that behave, think, and make decisions like humans.

AI Applications

• Healthcare
• Education
• Social Media
• Tourism
• Business
• Autonomous Vehicles
• Improving the world

Defining AI Behavior

• AI is technology that can learn and produce intelligent behavior.
• AI takes an input like pixels and processes it to output for example "Tuberculosis" diagnosis using computer vision.
• AI utilizes inputs like pixels for image to output "Four kids are playing with a ball" using computer vision.
• AI uses audio clip inputs like "I feel some eye pain" using speech recognition.

AI Features

• Object Detection
• Activity Recognition
• Semantic Segmentation
• Disease Detection
• Image Colorization
• Style Transfer
• Lip Sync
• Image to Image Translation

Reasons for Interest in AI

• Labor
• Science
• Appliances
• Search Engines
• Medicine/Diagnosis

AI Applications in Action

• A driverless car called STANLEY navigated the Mojave desert at 22mph and won the DARPA Grand Challenge.
• United Airlines uses automated speech recognition for flight bookings.
• NASA's REMOTE AGENT controlled spacecraft operations, planning, diagnosing, and recovering from issues.
• IBM's DEEP BLUE defeated Garry Kasparov in chess.
• Learning algorithms classify over a billion messages daily as spam.
• DART optimizes logistics planning and transportation scheduling.
• iRobot Corporation has sold over 2 million Roomba robotic vacuum cleaners.
• A computer program translates from Arabic to English using statistical models.

Branches of AI

• Machine learning
• Expert systems
• Deep learning
• Robotics
• Natural language processing
  • Speech recognition
  • Text to speech
  • Speech to text
  • Machine Translation
  • Vision
  • Image recognition
  • Text generation
  • Question answering
  • Classification
  • Context extraction

AI, ML, DL & Computer Science

• AI is a broad field encompassing machine learning and deep learning.
• Machine learning is a subset of AI.
• Deep learning is a subfield of machine learning.
• Computer science provides the foundation for AI.
• Mathematics, physics, chemistry, and biology provide input for Computer Science, Artificial Intelligence, Machine Learning, and Deep Learning.

AI in Real Time Systems

• Search engines, like Google Search, use AI for online ads.
• Recommendation systems by Netflix, YouTube, and Amazon use AI.
• AI drives Internet traffic through targeted advertising using AdSense and Facebook.
• Virtual assistants like Siri and Alexa use AI.
• Autonomous vehicles, like drones and self-driving cars use AI.
• AI enables automatic language translation with Microsoft Translator and Google Translate.
• Facial recognition on Apple Face ID or Microsoft DeepFace powered by AI.
• AI is used for image labeling on Facebook, Apple iPhoto, and TikTok, and for spam filtering.

Intelligent Systems

• Intelligent systems can be categorized based on thinking and behaving like humans versus rationally.
• Systems that think like humans
• Systems that think rationally
• Systems that behave like humans
• Systems that behave rationally

AI Approaches

• Cognitive Science Approach: Systems that think like humans
• Laws of Thought Approach: Systems that think rationally
• Turing Test Approach: Systems that behave like humans
• Rational Agent Approach: Systems that behave rationally

Cognitive Science Approach

• Requires human cognition model permitting simulation by computers.
• Deals with the reasoning processes (not only the behaviour).
• Aims to produce a human-like reasoning sequence in task-solving.
• Determining how humans think involves introspection, psychological experiments, and brain imaging.

Turing Test Approach

• Create machines that perform functions requiring intelligence when performed by people.
• Focuses on action, rather than representation of the world.
• A person, a computer, and an interrogator separated into three different rooms.
• The interrogator communicates with the two others only by teletype to avoid the machine imitating the appearance of voice of the person being tested
• If the machine fools the interrogator, it's deemed intelligent.

Laws of Thought Approach

• Study mental faculties with computational models to make it possible to reason and act.
• Focus is on inference mechanisms with guaranteed optimal solutions.
• The goal is to formalize reasoning as logical rules, procedures of inference.
• Systems allow inferences using "Socrates is a man. All men are mortal. Therefore Socrates is mortal".
• It is more general than using logic only because it uses logic and Domain knowledge.
• It allows extension of the approach with more scientific methodologies.

Rational Agent Approach

• Aims to emulate intelligent behavior through computational processes, automating intelligence.
• It focuses on acting sufficiently, and not optimally, in situations.
• The goal is to develop rational and sufficient systems.
• Obstacles include a need for 100% knowledge and too many computations.

Agents and Environments

• An AI system focuses studying rational agents and environments.
• The environment is sensed via sensors and agents act via actuators.
• AI agents exhibit mental properties, such as knowledge, belief, and intention.
• An agent perceives the environment through sensors and acts through actuators, following a cycle of perceiving, thinking, and acting.
• A sensor detects environmental changes, and agents observe through sensors.
• Actuators convert energy into motion, controlling the system. Actuators include electric motors, gears, etc.
• Effectors affect the environment. Effectors include legs, wheels, arms, and display screens.

Types of Agents

• Human Agents: sensory organs are sensory organs (eyes and ears) parallel the sensors and hands, legs, and the mouth act as effectors
• Robotic Agents: cameras and infrared act replace the sensors and various motors act and actuators act as effectors
• Software Agents: Keystrokes and files act as sensors and display on the screen, files, send network packets act as actuators

Intelligent vs Rational Agents

• Intelligent agents are autonomous entities employing sensors and actuators to achieve goals and learn from the environment.
  • Rule 1: Perceive the environment.
  • Rule 2: Use observation in decisions.
  • Rule 3: Decisions result in action.
  • Rule 4: AI action must be rational.
• Rational agents act to maximize performance with preferences and models for uncertainty.
• AI creates rational agents for game and decision theory in realistic scenarios.
• The rational action in an AI agent is critical for reinforcement learning, rewarding positive actions and penalizing negative ones.

Rational Agents

• It should do the right thing, based on its perceptions and available actions.
• The correct performance leads to success.
• It considers win/loss percentage for the overall maximize, robustness, and unpredictability to confuse the opponent.

Rational Agent Performance Measures

• Objective criteria define the success of an agent's behavior.
• Vacuum-cleaner performance: dirt cleaned, time taken, electricity used, noise generated.
• Self-driving car performance: time to destination, safety, predictability, reliability.
• Game-playing performance: win/loss percentage, robustness, unpredictability.

Rational Agents & Expectations

• For all percept sequences, it acts to maximize performance measurement based on evidence.
• This captures uncertain stochastic actions, used in stochastic environments to calculate action.
• It may also limit the worst-case behavior in high-risk settings.

Rationality

• Rationality differs from omniscience; rational behavior is possible with incomplete information.
• Agents gather information and explore to modify future percepts and an agent is autonomous if behavior stems from their own experiences.
• It can be judged on its performance measure defining success.
• It can require agent's prior knowledge of the environment, all actions available, and the sequence of percepts.

Structure of an AI Agent

• AI designs an agent program that implements an agent function based around:
  • Agent Architecture + Agent program
• Architecture is the machinery that an AI agent works off of.
• Agent Function maps a percept to an action via f:P* → A, where A is an action.
• Agent program implements Agent Function and operates of the physical architecture.

Agent Terminology

• Performance Measure: the criteria determining agent success.
• Behavior: The action taken after a given sequence of percepts.
• Percept : Agent's perceptual inputs.
• Percept Sequence: All agent perceptions to-date.
• Agent Function: Maps percept sequences to actions.
• Agent Program: An implementation of an agent function, acting within a system.
• Note: The agent function is an abstract mathematical description; and the agent program is a concrete implementation which operates withing a system.
• Tabulation
• It is in principle able to constructs this table by trying out all possible percept sequences and recording the actions the agent does in response.
• The table is an external characterization of the agent.
• The Agent function, is implemented by all agents in that program.

PEAS Model

• PEAS is a framework to define AI and rational agent properties.
• PEAS is measured by performance, environment, actuators, and sensors.
• Performance measures the success of an agent.

PEAS Examples for Self-Driving Cars

• Performance measures the safety, time, legal driving, and comfort.
• Environment involves roads, vehicles, signs, and pedestrians.
• The actuators use steering, acceleration, braking, signals, and horn.
• Sensors include camera, GPS, speedometer, odometer, accelerometer, and sonar.

PEAS Examples for Medical Diagnosis

• Performance: Healthy patient, minimized cost.
• Environment: Hospital, doctors, and patients.
• Actuator: Prescription, Diagnosis, and Scan report.
• Sensor: Symptoms and Patient's response.

PEAS Examples for Subject/Tutoring

• Performance: Maximize scores.
• Environment: Classroom, chair, desk, staff, and the students.
• Actuator: Smart displays, Corrections.
• Sensor: Eyes, Ears, and a notebook.

PEAS Examples for Vacuum Cleaner

• Performance: Cleanness, Efficiency, Battery Life, and Security.
• Environment: Desk, chair, the students, room, table, and the wood floor.
• Actuator: The wheels and Vacuum Extractor.
• Sensor: Camera, Dirt Detector, and Infrared wall.

Environment Types

• Fully vs Partially Observable:
  • Fully Observable: The agent can sense or access the environment's state at each point of time
  • Partially Observable: The opposite of full, which is when the sensors cannot sense and access everything.
  • An agent with no sensors in all environments is unobservable.
    • An example is when you play card games. Chess is Fully Observable
• Deterministic vs. Stochastic
  • Determinist is when selected actions can determine the next state of the environment.
  • Stochastic is when the environment is the opposite of a deterministic environment, in that the next state is unpredictable. An example is randomess.
    • An example for Deterministic would be traffic signals, and stochastic environment when the listen is not aware of the next song.
• Episodic vs. sequential is when the agent's actions divided into incidents or episodes.
  • The counterpart then depends on the action and what they are supposed to do in the future.
    • Pick and Place Robot and Tennis
• Static vs. Dynamic is based around wether an agent is acting, if so, it is static. Therefore its counterpart can change during an agent's actions.
  • Cleaning and Soccer
• Discrete vs. Continuous
  • An environment with finite number of states and agents taking finite number of actions.
  • The counterpart is an environment with infinite number of states. Therefore the possibilities of taking action are infinite.
    • Tic-Tac-Toe and Basketball Game
• Single vs. Multi-agent
  • An environment that is called single, then only one can operate as an environment that would be a single agent.
  • Multi can have numerous depending on different problems from similar environments.
    • Maze and Football
• Competitive vs Collaborative depends on optimizing the output based on other agents to get their desired output.
  • Competition: Each agent plays each other to win
  • Collaboration: Cooperate to have a desired output
    • Chess and Self-Driving Cars

Types of Agents

• Simple reflex agents use the basis of a percept and do not based it on other perceptions
• This function is based off condition-action where the state is mapped to an action, and used when true.
• For an easier function, its environment should be fully-observable in order to be escaped on its random actions.
• With this then the vacuum agent becomes a simple agent where its current location is based on its cleanliness.

Model-based reflex agents

• Can work in a partially observable environment and track the situation it is in.
• Based on model-based has two important factors which are:
  • Agent State requires updates and must contain information:
    • How the world evolves.
    • How agents have actions that would affects the world.

Goal-based agents

• Goal-based agents expand off model-based by stating additional goal information to do.
• The knowledge of the current state can be not sufficient in an environment to find a choice of what to do.
• Agents choose because of the achieve goals that are based off the current state.
• These long-term goals are used in a long sequence of possible actions to measure if a goal is achievable.
• When comparing agent, the automated taxi reflex agent is the one that applies brakes when it sees the lights, where the goal is to stop others.

Model-Based Reflex Agents

• Model-Based Reflex Agents has knowledge about "how things happen in the world" , which is call it that.
• The importance is when the current state is being read from all the percepts.
• It requires for the Agent to know the Updates and Information for:
  • The World Evolving and how it can apply
  • How would the agents action would end
  • In short it can work with the least amount of environment and the most of percept's history.

Utility-based agents

• Utility is the end use that based the building blocks of the action and if not can be avoided by alternatives.
• Instead of just doing only to achieve the best goal, now can it come in quicker when compared.
• Happiness to the maximum amount can be taken into account and considered as a whole.
• It is efficiency to the goal, so when taken into account should the end goal have an action/method to get there quickly, instead.
• A number that with the help of the agent function mapping the state towards in order to get the goal.

Learning element in an agent

• It makes improvements by learning from the environment.
• The critic provides feedback to the agents to measures the agent's performance.
• Performance element makes select the external action.
• The agent is being told to find new innovative experiences.
• Learning agent is able to adapt and act by the usage of past learnt actions.

Agent Summary

• An agent can act and learn from itself while using an programmed architecture.
• Action maximize based on the current past and can take future expectations into account.
• If can build around itself to act it can get past limited settings due to environment.

Agent Types

• Table-driven: Uses look-up table implemented into the memory as it is based off the past
• Simple Reflex: Can be seen to apply rules in order to achieve a pre-made state or action rule in that environment.
• With Memory: A form agent with its used to keep internal tracking of the past states that as gone since then.
• With Goals: Goal-baed helps that the agent not only considers as past experiences and actions, but ones that help meet to the end point goal.

Utility-based Agents

• Base their decisions on traditional utility concepts and in it have action based rationally, as they consider all angles.
• Not every agent only focuses on its environment, but can get better or as time goes by can find itself improving.
• In what ever the environment shows it to be, the agent will end up with actions better than the human counterpart in these kinds of examples.