Unit 1: Foundations and AI Problem Solving

I. Orientation

Artificial Intelligence (AI) is the branch of computing concerned with constructing systems that perform tasks normally associated with human intelligence. The field emerged as a formal discipline at the Dartmouth workshop (1956), although its foundations include earlier work in logic, mathematics, psychology and computation.

A. Introduction to Artificial Intelligence

AI enables an artificial agent to perceive its environment, reason about available information and act toward a goal.

• Intelligence: The capacity to learn, reason, solve problems, understand language, recognize patterns and adapt to change.
• Artificial agent: A software or robotic entity that receives percepts through sensors and affects its environment through actuators.
• Rationality: Selection of the action expected to maximize a specified performance measure, given available evidence and computational limits.
• Core capabilities:
  • Perception: Interpreting images, speech or sensor readings.
  • Knowledge representation: Encoding facts and relationships through rules, graphs or vectors.
  • Reasoning: Drawing conclusions from represented knowledge.
  • Learning: Improving behavior using data or experience.
  • Planning: Choosing an action sequence that reaches a goal.
• Interdisciplinary foundations: AI draws from computer science, mathematics, statistics, neuroscience, linguistics, philosophy and control theory.
• Central principle: Intelligence is treated computationally by defining inputs, internal representations, goals, actions and measurable outcomes.

II. Historical Development

A. History and evolution of Artificial Intelligence

AI developed through alternating periods of theoretical progress, practical success, excessive expectations and reduced funding.

• Early foundations, 1940s-1950s: McCulloch and Pitts proposed a mathematical neuron model in 1943; Alan Turing introduced the imitation-game criterion in 1950.
• Birth of the field, 1956: The Dartmouth workshop, organized by researchers including John McCarthy and Marvin Minsky, established “Artificial Intelligence” as a research identity.
• Symbolic AI, 1950s-1960s: Programs used logic, symbols and search; examples include the Logic Theorist (1956) and General Problem Solver (1957).
• First AI winter, 1970s: Limited computing power, poor scalability and unmet promises caused funding and confidence to decline.
• Expert systems, 1970s-1980s: Systems such as MYCIN encoded specialist knowledge as rules of the form IF condition THEN conclusion.
• Second AI winter, late 1980s-1990s: Expert systems proved expensive to maintain and often failed outside narrowly defined conditions.
• Statistical learning, 1990s-2000s: Probabilistic models, support-vector machines and data-driven evaluation became prominent; IBM Deep Blue defeated Garry Kasparov in 1997.
• Deep-learning era, 2010s: Larger datasets, graphics processors and multilayer neural networks produced major advances; AlexNet achieved influential ImageNet results in 2012.
• Generative AI, 2020s: Transformer-based foundation models began generating coherent text, images, audio and code from natural-language instructions.

III. Scope and Classification

A. Characteristics and types of Artificial Intelligence

AI systems are classified by both the capabilities they exhibit and the way they operate.

• Characteristics: Effective systems show autonomy, goal-directed behavior, adaptability, inference under uncertainty and interaction with an environment.
• Narrow AI: Performs a restricted task, such as spam filtering or face recognition; nearly all deployed AI belongs to this category.
• Artificial General Intelligence: A proposed system able to transfer knowledge and perform intellectual tasks across many domains at human-level breadth.
• Artificial Superintelligence: A hypothetical intelligence exceeding human performance across most cognitive activities.
• Reactive machines: Respond only to current inputs without maintaining an internal history; a fixed game-playing policy illustrates this type.
• Limited-memory systems: Use recent or learned historical information; autonomous-driving models process past observations and current sensor data.
• Theory-of-mind AI: A research-stage concept involving representation of beliefs, intentions and emotions of other agents.
• Self-aware AI: A hypothetical system possessing consciousness or an explicit subjective understanding of itself.
• Knowledge-based versus learning-based AI: The former applies manually represented rules; the latter estimates patterns from examples.

IV. Related Computational Fields

A. Artificial Intelligence vs Machine Learning vs Deep Learning vs Data Science

These fields overlap, but they differ in scope, methods and intended outputs.

1. Artificial Intelligence: The broad objective of creating intelligent behavior; it includes search, planning, knowledge-based systems, robotics and learning.
2. Machine Learning: A subset of AI in which algorithms improve task performance from data rather than relying entirely on explicit rules.
3. Deep Learning: A subset of machine learning using neural networks with multiple representation layers; convolutional and transformer networks are examples.
4. Data Science: An interdisciplinary practice for extracting knowledge from data through statistics, programming, visualization and domain analysis; its output may be insight rather than an autonomous agent.

• Set relationship:

Deep Learning ⊂ Machine Learning ⊂ Artificial Intelligence
Data Science overlaps with all three but is not wholly contained in AI.

• Concrete contrast: A rule-based chess program is AI but not necessarily machine learning; a sales dashboard can be data science without being AI.
• Typical goals: AI chooses intelligent actions, machine learning predicts or discovers patterns, deep learning learns complex representations, and data science supports evidence-based decisions.

V. Intelligent-System Development

A. Intelligent systems and Artificial Intelligence lifecycle

An intelligent system integrates sensing, reasoning or learning, and action, while the AI lifecycle organizes its development from problem definition to continuing operation.

• System components:
  • Sensors or inputs: Cameras, microphones, databases or user requests.
  • Model or knowledge base: Learned parameters, facts, rules or world representations.
  • Inference mechanism: Produces predictions, plans or decisions.
  • Actuators or outputs: Displays, recommendations, control signals or robotic motion.
• Problem definition: Identify stakeholders, goals, constraints and a measurable target; for example, predict equipment failure at least 24 hours in advance.
• Data acquisition: Gather representative, lawful and sufficiently reliable observations from sensors, records or experiments.
• Preparation: Clean missing values, label examples, engineer features and divide data into training, validation and test sets.
• Model development: Select an approach, train it on the training set and tune settings against validation performance.
• Evaluation and deployment: Test accuracy, robustness, fairness, latency and safety before integrating the model into real workflows.
• Monitoring and improvement: Detect data drift, performance decay and harmful outcomes; retrain, revise or retire the model when required.
• Governance throughout: Privacy, security, explainability, accountability and human oversight apply across every lifecycle stage.

VI. Sectoral Uses

A. Applications of Artificial Intelligence in healthcare, agriculture, finance, education, manufacturing, robotics and smart cities

AI creates value by converting data into predictions, decisions or automated actions, although each sector requires domain-specific safeguards.

• Healthcare: Medical-image analysis can locate suspicious lesions; clinical decision support and drug discovery assist professionals, but diagnosis requires validation, privacy protection and human supervision.
• Agriculture: Satellite imagery, soil sensors and weather data support crop-disease detection, yield forecasting, precision irrigation and targeted pesticide application.
• Finance: Models detect fraudulent transactions, estimate credit risk, automate document processing and support algorithmic trading; biased training data can produce discriminatory decisions.
• Education: Adaptive platforms vary content according to learner performance, while automated feedback and learning analytics identify areas requiring teacher intervention.
• Manufacturing: Predictive-maintenance models estimate failure from vibration or temperature readings; computer vision detects defects and robots automate repetitive assembly.
• Robotics: AI combines perception, localization, planning and control so robots can navigate, manipulate objects and collaborate with humans.
• Smart cities: Traffic-signal optimization, energy-demand forecasting, waste routing and infrastructure monitoring can improve resource use; surveillance applications raise significant privacy concerns.

VII. Formalizing AI Tasks

A. Artificial Intelligence problem formulation and state-space representation

Problem formulation converts a real-world objective into a precise model that a search or reasoning procedure can process.

• Initial state: The situation before any action; in route finding, it is the starting city.
• State: A sufficient representation of a possible situation, such as (robot_position, battery_level).
• Actions: Operations available in a state, written as Actions(s).
• Transition model: The result of applying action a in state s:

Result(s, a) = s'

Here, s is the current state, a is an action and s' is the successor state.

• Goal test: A predicate determining whether a state satisfies the objective, such as position = destination.
• Path cost: The accumulated cost of actions:

g(n) = Σ c(si, ai, si+1)

Here, g(n) is the cost to node n, and c is the step cost between successive states.

• State space: The set of states reachable from the initial state through valid actions.
• Abstraction: Irrelevant details are omitted; route planning may represent cities and roads while ignoring building colors and vehicle upholstery.

VIII. Evaluating Problems and Outcomes

A. Problem characteristics

Problem characteristics determine which representation and solution method are appropriate.

• Observability: A fully observable problem reveals the complete state; a partially observable one provides incomplete or noisy evidence.
• Determinism: Deterministic actions have one predictable result, while stochastic actions have probabilistic outcomes.
• Structure: Problems may be discrete or continuous, static or dynamic, episodic or sequential, and single-agent or multi-agent.
• Knowledge status: In known environments, transition rules are available; in unknown environments, an agent must learn them.
• Decomposability: Some problems divide into independent subproblems, whereas tightly coupled problems require joint optimization.
• Reversibility: Chess moves are generally irreversible, while many puzzle moves can be undone; reversibility affects recovery from poor choices.

B. Search space and solution space

The search space contains all candidate states considered by a procedure, whereas the solution space contains candidates that satisfy the goal conditions.

1. Search space: Represented as a graph whose nodes are states and edges are actions; cycles and repeated states may greatly increase exploration.
2. Solution space: Contains valid goal states or valid paths; an optimal-solution subset contains those with minimum cost or maximum utility.

• Branching factor: If each state has approximately b successors and a shallowest solution lies at depth d, breadth-first exploration can require approximately O(b^d) nodes.
• Constraint effect: Legal-move rules, capacity limits or scheduling requirements eliminate invalid candidates before or during search.
• Optimization: A feasible solution satisfies all constraints; an optimal solution also minimizes or maximizes a stated objective.

C. Performance measures

A performance measure specifies how success is judged and prevents “intelligent” behavior from being defined only by appearance.

• Solution quality: Measured by cost, utility, reward, accuracy or distance from an optimum.
• Completeness: Whether an algorithm is guaranteed to find a solution when one exists.
• Optimality: Whether the returned solution is guaranteed to have the best objective value.
• Time complexity: The number of operations or generated nodes as problem size grows.
• Space complexity: The maximum memory required during execution.
• Predictive metrics: Classification commonly uses accuracy, precision, recall and F1-score:

Precision = TP / (TP + FP)
Recall = TP / (TP + FN)
F1 = 2 × Precision × Recall / (Precision + Recall)

Here, TP, FP and FN mean true positives, false positives and false negatives.

• Operational measures: Latency, energy use, safety, robustness, fairness and human satisfaction may matter more than raw accuracy.

IX. Problem-Solving Methods

A. Overview of Artificial Intelligence-based problem-solving approaches

AI problem solving selects a method according to available knowledge, uncertainty, scale, feedback and required guarantees.

• Uninformed search: Breadth-first, depth-first and uniform-cost search use only the problem definition; they differ in completeness, memory use and cost handling.
• Informed search: Greedy best-first search and A use a heuristic h(n) estimating remaining cost; A evaluates f(n) = g(n) + h(n).
• Constraint satisfaction: Variables, domains and constraints represent scheduling, assignment and configuration tasks; backtracking systematically tests consistent assignments.
• Adversarial search: Minimax models competing agents in games, while alpha-beta pruning removes branches that cannot affect the final decision.
• Knowledge-based reasoning: Logical rules and ontologies support deduction; probabilistic models such as Bayesian networks reason under uncertainty.
• Planning: A planner constructs action sequences from an initial state to a goal while respecting action preconditions and effects.
• Optimization methods: Hill climbing, genetic algorithms and simulated annealing search large spaces where exact methods are impractical.
• Machine learning: Supervised, unsupervised and reinforcement learning derive predictive structures or policies from examples and feedback.
• Hybrid approaches: Practical systems often combine learned perception with symbolic rules, search, optimization and human review.