Artificial Intelligence Foundations

Artificial Intelligence Foundations

• Artificial intelligence (AI) studies the synthesis and computational agents that act intelligently. AI aims to create computers and programs capable of intelligent action.

What is AI?

• AI is the theory and development of computer systems that perform tasks typically requiring human intelligence.

Examples of Intelligence

• Logical reasoning
• Problem-solving
• Creativity
• Planning

Narrow vs. General AI

• Narrow AI requires reconfiguration or a new algorithm for each new task (e.g., chess, speech recognition, facial recognition).

• General AI applies intelligent systems to any problem, applying knowledge across a range of tasks.

• The primary research areas of AI are reasoning, learning, problem-solving, and perception.

Applications of AI

• Robotics (industrial, autonomous, domestic)
• Industrial automation (intelligent control, safety, security)
• Healthcare (drug design, operating room robotics)
• Games (NPCs, virtual, augmented reality)
• Other areas (education, agriculture, personal assistance)

Intelligent Agents

• An agent is anything perceiving its environment through sensors and acting upon it using actuators.
• Intelligent agents perceive, reason, and act autonomously to achieve goals.
• The percept sequence is the complete history of all data received by an agent from its sensors.

Agent Rationality

• A rational agent acts to achieve the best possible outcome based on its knowledge and goals, aiming to maximize its performance.
• Prior knowledge, performable actions, percept sequence to date, and success criteria are important factors.
• Observability is crucial, and partially observable environments present challenges.

Discrete vs. Continuous Environments

• Discrete environments have a finite number of actions and states (e.g., chess).
• Continuous environments have infinite possibilities of actions or states (e.g., tennis).
• Adversarial environments involve competition between agents, while benign environments do not.

Types of Agents

• Reflex agents use the current percept.
• Model-based reflex agents keep track of environment state.
• Goal-based agents plan actions to achieve goals.
• Utility-based agents maximize utility.
• Learning agents adapt over time.

Search Techniques

• Search is a powerful technique for solving AI problems, often formulated as directed graphs.
• The goal is to find the correct action sequence to achieve the goal while respecting constraints.
• Blind search explores without prior knowledge.
• Informed search uses heuristic functions to guide the search (e.g., Breadth-First Search, Depth-First Search).

Modelling Challenges

• Algorithmic complexity measures the computational resources needed by an algorithm based on the input size (often represented using Big O notation).

Knowledge and Reasoning

• Knowledge in AI refers to the data, information, and concepts used by AI systems to understand the world and solve problems.
• Reasoning is the process of drawing conclusions using available knowledge.

Knowledge-Based Agents

• A knowledge-based agent uses a knowledge base to make decisions, reason, and solve problems, maintaining and updating its knowledge base.

Knowledge Representation

• A knowledge base contains statements (sentences) in a representation language.
• Axioms are sentences not derived from other entities.
• TELL operations add new sentences.
• ASK operations query the knowledge base.

First Order Logic (FOL)

• A formal language for representing objects, functions, and relationships in an environment.
• FOL models include objects, constants, functions, and relations to describe the environment.

Logical Inference

• Logical inference is the process of drawing conclusions based on some known information and reasoning.
• Desirable properties for inference algorithms include soundness (deriving only entailed conclusions) and completeness (deriving all entailments).

Planning

• Planning is the process of creating a strategy to achieve a desired goal state by choosing actions that maximize the probability of success.
• Planning frequently involves identifying tasks, actions, and their ordering.
• Planning differs from scheduling in that planning focuses on identifying actions, while scheduling focuses on the timing of those actions.

AI Planning Techniques

• Domain-specific planning is tailored to a particular domain (e.g., a game).
• Domain-independent planning uses techniques applicable to various domains.
• Planning techniques include state-space search and plan-space search to find solutions.

Satisficing vs. Optimal Planning

• Satisficing planning aims to find a good enough solution, often prioritizing speed.
• Optimal planning aims to find the best possible solution under specific criteria (e.g., minimum steps, lowest cost).

Contingent Planning

• Addresses uncertainty by considering different possible outcomes of actions. The plans also account for how actions could affect future outcomes.

Conformant Planning

• Plans that work regardless of the uncertain outcome.
• The agent will plan without knowing precisely how the environment or future will unfold.

Temporal Planning

• Includes planning with constraints
• Includes numeric resources like fuel for travel

Uncertainty and Probability in AI

• Stochastic events and partial observability introduce uncertainty into AI systems.
• Probability distributions help model uncertainty, enabling a more robust estimate of future outcomes given current conditions.
• Bayes Networks are representations for probability distributions.

Time, Resources, and Exogenous Events

• Real-world planning often involves temporal constraints, resource limitations, and unforeseen events that are not predictable.

Machine Learning

• A sub-field of AI focused on developing models from data.

Classification vs. Regression

• Classification aims to categorize data into discrete classes.
• Regression aims to predict continuous values.

Linear Regression

• A simple, widely used algorithm for regression.
• Attempts to fit a straight line through data points to predict values.

Reinforcement Learning

• A machine learning method for training agents to make decisions in an environment to maximize rewards. The agent learns through trial and error based on feedback responses.

Active and Passive Reinforcement Learning

• Passive reinforcement learning involves learning a value function given a pre-determined policy, providing a method for evaluating actions.
• Active RL agents modify their environment to learn about it and improve their decision-making.

Markov Decision Processes (MDP)

• A formal framework for modelling sequential decision-making problems with uncertainty.

Exploration-Exploitation Trade-Off

• Exploration involves trying unseen actions/paths to maximize potential gain, while exploitation involves prioritizing actions already known to enhance rewards.
• The balance between exploration and exploitation must be considered to achieve long-term, optimized performance.

Computer Vision

• Extracts information, analyzes, and interprets visual data (images/videos).
• Used in object recognition, image processing, analysis, detection.

Thresholding

• A simple way to classify or make decisions on data by comparing values against a fixed value.
• Widely used in image processing and anomaly detection.

Law and Ethics in AI

• The ethical considerations in AI systems including bias, transparency, accountability, data privacy.

Natural Language Processing (NLP)

• A branch of AI focused on enabling computers to understand, interpret, and generate human language.

Semantic Networks

• Represents knowledge using interconnected nodes and edges to show relationships.
• Traversing nodes in this network, depending on the edge type, determines information relationships and associations.

Description Logics

• A formal language to build and link categories.

Online State Estimation

• Algorithms predicting the most likely state of systems from real-time data.
• Particle filters estimate the belief state, updating probabilistic estimations over time.

State Representation

• Models describing the state of a system.
• Includes kinematic (position, velocity) as well as dynamic representations of the state.

Alternative Systems for Knowledge Representation

• Semantic networks use nodes and edges to describe relationships.
• Description logics provide a formal means for constructing and combining categories and their relationships.

Multi-Agent Systems

• Systems where more than one agent interacts with each other.

Zero-Sum Games

• Games where one player's gain precisely equals another player's loss.

Minimax Algorithm

• Often applied to games of strategy.
• Determines the best move for a player under the assumption that the opponent will play the worst possible move.

Alpha-Beta Pruning

• A method to optimize the minimax search algorithm, reducing computational load.

Game Theory

• Game theory examines strategic interactions between multiple entities (players).

Explore the fundamentals of artificial intelligence, including its definition, examples of intelligence, and the distinction between narrow and general AI. This quiz covers the key applications and research areas of AI, such as robotics and healthcare.