Foundations of Artificial Intelligence

Introduction

Artificial intelligence (AI) is the study and engineering of systems that perform tasks normally associated with human intelligence. These tasks include perceiving the environment, understanding language, learning from experience, solving problems, making decisions, and acting toward goals. AI is not a single technology. It is a broad field that combines ideas from computer science, mathematics, statistics, psychology, linguistics, neuroscience, philosophy, and engineering.

AI has become part of everyday life. Search engines rank results, recommendation systems suggest films and products, navigation applications estimate travel times, and fraud-detection systems identify unusual transactions. However, the existence of intelligent-looking software does not mean that machines think in exactly the same way as humans. Understanding the foundations of AI requires distinguishing different types of intelligence, examining how AI systems operate, and recognizing both their capabilities and limitations.

Learning Objectives

By the end of this chapter, you should be able to:

• define artificial intelligence and distinguish it from conventional software;
• differentiate narrow AI, artificial general intelligence, and superintelligence;
• describe major historical developments in AI;
• explain intelligent agents, rationality, task environments, and agent architectures;
• distinguish fully observable, partially observable, deterministic, stochastic, static, dynamic, discrete, and continuous environments;
• explain the roles of knowledge representation, search, and reasoning;
• identify common AI applications and their professional implications;
• evaluate practical, ethical, and governance challenges in AI projects.

1. What Is Artificial Intelligence?

There is no single definition accepted by every researcher. Definitions differ according to whether intelligence is understood as human-like behavior, human-like thought, logical reasoning, or successful goal-directed action. A useful engineering definition is:

AI is the design of computational agents that perceive or receive information about an environment, process knowledge, learn or reason, and select actions intended to achieve goals.

An AI system generally has four related capabilities:

1. Perception or input processing: receiving data from sensors, databases, documents, images, audio, or users.
2. Representation: organizing information into forms that a program can manipulate.
3. Inference, learning, or planning: deriving conclusions, identifying patterns, predicting outcomes, or selecting a course of action.
4. Action or output: producing a recommendation, classification, generated response, control signal, or physical movement.

Conventional software usually follows explicitly programmed rules for anticipated situations. AI systems may also use rules, but many modern systems learn relationships from examples. For instance, a traditional program may contain a rule stating that an order above a specified value requires approval. A machine-learning fraud detector may instead estimate whether a transaction is suspicious from historical transactions and their outcomes.

The boundary is not absolute. Many practical systems combine deterministic programming, databases, statistical models, and machine learning. AI should therefore be understood as a collection of methods for building goal-directed, adaptive, or knowledge-based computational systems rather than as a magical substitute for human judgment.

2. Narrow AI and General AI

Narrow or Weak AI

Narrow AI is designed for a limited task or a closely related group of tasks. A speech-recognition model can convert spoken language into text, but it does not necessarily understand the speaker’s intentions. A chess program can search chess positions effectively, but it cannot independently diagnose illness or repair a machine.

Narrow AI can be highly capable within its specified domain. Its limitation is not necessarily low performance; it is restricted scope. The system’s knowledge, objectives, inputs, and outputs are bounded by its design and training.

Artificial General Intelligence

Artificial general intelligence (AGI) refers to a hypothetical system with flexible, broad capabilities comparable to human general intelligence. Such a system would be expected to transfer knowledge between domains, learn new tasks efficiently, reason under unfamiliar conditions, and adapt to changing goals and environments.

AGI has not been scientifically established as an achieved technology. Claims about it should therefore be distinguished from demonstrated performance on particular benchmarks or applications.

Superintelligence

Artificial superintelligence is a hypothetical form of intelligence that would substantially exceed human cognitive performance across most or all relevant domains. It remains a subject of forecasting, philosophy, safety research, and public debate rather than an established engineering category.

3. Historical Development of AI

AI developed through several overlapping traditions rather than through one continuous sequence of inventions.

Foundations before 1950

Formal logic provided methods for representing propositions and deriving conclusions. Probability and statistics supplied tools for reasoning under uncertainty. In 1943, Warren McCulloch and Walter Pitts proposed a mathematical model of artificial neurons, demonstrating how simplified units could represent logical operations. In 1950, Alan Turing published “Computing Machinery and Intelligence,” which asked whether machines could think and proposed an imitation-game approach—later called the Turing test—as one way to discuss machine intelligence.

The birth of the field

The term artificial intelligence is strongly associated with the 1956 Dartmouth workshop, organized by John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon. The workshop helped establish AI as a research field and encouraged work on symbolic reasoning, problem solving, language, and learning.

Symbolic AI and early programs

Early systems represented objects and rules explicitly. Programs such as the Logic Theorist and the General Problem Solver demonstrated automated reasoning in restricted settings. Researchers also developed game-playing programs, theorem provers, and early natural-language systems. This period is often called symbolic AI or “good old-fashioned AI.”

Expert systems

During the 1970s and 1980s, expert systems encoded specialist knowledge as rules. A medical expert system, for example, might combine symptoms with diagnostic rules. Expert systems showed that useful performance could be achieved in narrow domains, but they were expensive to construct and maintain. The process of collecting knowledge from experts was known as the knowledge-acquisition bottleneck.

AI winters

Periods of reduced funding and enthusiasm are known as AI winters. They occurred partly because early systems were less robust than their demonstrations suggested, computing resources were limited, and ambitious promises exceeded practical results. These periods provided an important lesson: technical claims should be evaluated against reproducible evidence and operational requirements.

Machine learning and statistical AI

From the 1990s onward, statistical methods and machine learning became increasingly influential. Instead of specifying every rule, developers trained models from data. Improvements in algorithms, digital data, hardware, and storage supported advances in speech recognition, recommendation, computer vision, and natural-language processing.

Deep learning and foundation models

Deep learning uses multi-layer neural networks to learn representations from large datasets. Major progress in image recognition, speech processing, and language modelling occurred during the 2010s. More recently, foundation models trained on broad datasets have been adapted for tasks such as text generation, image synthesis, coding assistance, and multimodal interaction. These systems are powerful but can still produce inaccurate outputs, inherit biases, expose sensitive information, or behave unpredictably outside their training conditions.

4. Intelligent Agents

An agent is an entity that perceives its environment through inputs and acts upon that environment through outputs. A human uses eyes, ears, and other senses as sensors and muscles as actuators. A robot may use cameras and lidar as sensors and motors as actuators. A software agent may receive web requests and database records and produce messages, rankings, or transactions.

An agent function maps a history of percepts to an action. In conceptual notation:

Here, P* is the set of possible percept histories and A is the set of available actions. The notation describes an abstract mapping; it does not prescribe a particular implementation. An agent program is the actual software that implements, approximates, or learns this mapping on a machine.

A useful agent cycle is:

Perceive → Represent → Decide → Act → Receive feedback

The cycle may occur in milliseconds for an autonomous vehicle or over hours for a planning system. Feedback is important because the consequences of actions can change later decisions.

Rational agents

A rational agent selects an action expected to maximize a performance measure, given its percept history, available knowledge, and computational resources. Rationality is not the same as perfection. An agent may make a rational decision with limited information and still obtain an undesirable result because the environment is uncertain.

A general expected-utility expression is:

Here, EU(a | e) is the expected utility of action a given evidence e; s represents possible outcomes or states; P(s | e, a) is the probability of outcome s given the evidence and action; and U(s) is the utility assigned to that outcome. The summation applies to discrete outcomes. For continuous outcomes, an integral is used instead. The expression assumes that probabilities and utilities are defined consistently and that the decision-maker is comparing actions using the same performance criterion.

For example, a delivery agent may choose between a fast but expensive route and a slower inexpensive route. The rational choice depends on delivery deadlines, fuel cost, traffic uncertainty, and the organization’s performance measure—not simply on distance.

Agent architectures

Common agent designs include:

• Simple reflex agents: choose actions from current percepts using condition–action rules. They work best in simple, fully observable environments.
• Model-based agents: maintain an internal model of aspects of the world that cannot be observed directly.
• Goal-based agents: select actions by considering whether they move the system toward a specified goal.
• Utility-based agents: compare possible outcomes using preferences or utilities, allowing trade-offs between competing objectives.
• Learning agents: improve performance through experience. A learning agent commonly includes a performance element, a learning element, a critic that evaluates outcomes, and a mechanism for proposing exploratory actions.

5. Task Environments

The behavior of an agent cannot be evaluated without describing its environment. A common framework specifies the task using PEAS:

• Performance measure: how success is evaluated;
• Environment: the external world in which the agent operates;
• Actuators: mechanisms through which it acts;
• Sensors: mechanisms through which it receives information.

For an automated taxi, the performance measure may include safety, legality, travel time, passenger comfort, and operating cost. The environment includes roads, vehicles, pedestrians, weather, and traffic rules. Actuators include steering, braking, and acceleration. Sensors include cameras, radar, lidar, maps, and positioning systems.

Important environment properties include:

This classification guides system design. A partially observable, stochastic, dynamic environment generally requires more than a simple rule table; it may require state estimation, prediction, planning, and risk management.

6. Knowledge Representation

Knowledge representation concerns how information about a domain is encoded so that a computer can use it. A representation should support relevant inference while remaining understandable, maintainable, and computationally manageable.

Important approaches include:

• Logic: represents facts and relationships using formal statements and inference rules.
• Production rules: represents knowledge as condition–action or condition–conclusion rules.
• Semantic networks and knowledge graphs: represent entities as nodes and relationships as labelled links.
• Frames: organize stereotyped objects or situations into structures with attributes and values.
• Ontologies: specify concepts, categories, properties, and relations in a domain.
• Vectors and embeddings: represent objects as numerical coordinates so that statistical models can identify patterns or similarities.
• Probabilistic representations: represent uncertainty through probabilities or probability distributions.

A representation involves trade-offs. A symbolic rule can be easy to inspect but difficult to create at large scale. A learned vector representation can capture complex patterns but may be difficult to interpret. Practical systems often combine both approaches, for example using machine learning to identify entities and a knowledge graph to organize validated relationships.

Knowledge may be explicit, such as a stored policy, or implicit, such as patterns encoded in model parameters. In both cases, data quality, coverage, consistency, and update procedures influence system reliability.

7. Search, Problem Solving, and Planning

Many AI problems can be expressed as search. A problem formulation typically includes an initial state, actions, a transition model describing the results of actions, a goal test, and a path-cost function.

For a route-planning problem, the initial state is the starting location, actions are legal road transitions, the transition model describes movement between locations, the goal test checks whether the destination has been reached, and the path cost may measure distance, time, or monetary expense.

Uninformed search

Uninformed methods do not use domain-specific estimates about how close a state is to the goal. Breadth-first search explores shallow paths first and can find a shortest path when every step has equal cost. Depth-first search explores one branch deeply before backtracking; it can use less memory but does not generally guarantee a shortest solution.

Informed search

Informed methods use a heuristic, written as h(n), to estimate the remaining cost from state n to a goal. A common evaluation function is:

Here, g(n) is the cost already incurred, h(n) is the estimated remaining cost, and f(n) is the priority assigned to the state. The A algorithm uses this evaluation. If the heuristic never overestimates the true remaining cost and costs are non-negative, A can return an optimal solution under its standard assumptions. If the heuristic is inaccurate, the algorithm may still find a solution but optimality or efficiency guarantees may change.

Planning and optimization

Planning extends search by selecting actions over time to achieve goals. In real systems, planning must account for uncertainty, changing conditions, resource limits, and conflicting objectives. A warehouse robot, for example, may need to minimize travel time while avoiding collisions, preserving battery power, and meeting delivery priorities.

8. Reasoning and Learning

Reasoning derives conclusions from available information. Deductive reasoning applies general rules to particular facts. Inductive reasoning generalizes from examples, although its conclusions may be uncertain. Abductive reasoning seeks the most plausible explanation for observations.

AI systems reason in several ways:

• logical inference from facts and rules;
• probabilistic inference under uncertainty;
• constraint satisfaction, where possible solutions must satisfy restrictions;
• causal reasoning, which examines how interventions may change outcomes;
• case-based reasoning, which adapts solutions from similar previous cases.

Machine learning enables a system to improve its performance from data or experience. In supervised learning, a model learns from examples containing inputs and target outputs. In unsupervised learning, it identifies structure without specified target labels. In reinforcement learning, an agent learns through interactions and rewards or penalties.

AI, machine learning, and deep learning are related but not identical. AI is the broad field. Machine learning is one approach within AI. Deep learning is a family of machine-learning methods based primarily on multi-layer neural networks.

9. Common Applications of AI

AI applications include:

• Natural-language processing: translation, search, summarization, speech recognition, and conversational systems.
• Computer vision: inspection, medical-image analysis, object detection, and document processing.
• Recommendation: ranking products, videos, articles, or learning resources.
• Healthcare: decision support, triage assistance, drug discovery, and workflow automation. Clinical use requires professional oversight and validation.
• Finance: fraud detection, credit-risk analysis, forecasting, and compliance monitoring.
• Manufacturing: predictive maintenance, quality inspection, process optimization, and robotics.
• Transportation: route planning, demand prediction, driver assistance, and logistics.
• Education: adaptive practice, feedback, accessibility tools, and learning analytics.
• Cybersecurity: anomaly detection, malware analysis, and alert prioritization.
• Public services: resource allocation, document classification, and emergency response support.

An application should not be judged merely by whether a model is accurate in a laboratory. Deployment also depends on reliability, latency, cost, security, interpretability, accessibility, legal compliance, and the consequences of error.

10. Illustrative Case: AI-Assisted Visual Quality Inspection

Background

A manufacturing company inspects packaged products on a conveyor. Previously, workers visually examined every item. The process was slow, physically repetitive, and inconsistent for subtle defects.

Problem

The company wants to identify damaged packaging and incorrect labels while maintaining production speed. A false negative allows a defective item to reach a customer. A false positive causes a good item to be removed for manual review.

Analysis

The task is a narrow-AI classification problem. Cameras provide sensor input; conveyor controls and a rejection mechanism are actuators. The environment is dynamic, partially observable, and subject to variation in lighting, product orientation, and packaging design. The performance measure must combine detection quality, inspection speed, equipment cost, and the relative consequences of the two error types.

Approach and solution

The company collects representative images, including normal products and several defect categories. Quality specialists label the images and define acceptance criteria. A vision model is trained and evaluated on data kept separate from training data. Lighting and camera position are standardized where possible. The system sends uncertain cases to a human inspector rather than forcing every case into an automatic decision.

The deployment also includes monitoring for changes in product design, camera quality, and defect frequency. Access controls protect images and production data. Operators can override the system, and override rates are reviewed as evidence of possible model deterioration.

Outcome

The system improves throughput and provides consistent first-pass inspection, but it does not eliminate human responsibility. Human review remains necessary for ambiguous cases, unusual defects, model updates, and safety decisions.

Learning points

1. A clear operational objective is more useful than a vague goal such as “use AI.”
2. The environment and error costs determine the appropriate agent design.
3. Data quality and representative coverage are as important as model selection.
4. Deployment requires monitoring, fallback procedures, and accountability.
5. Narrow AI can create substantial value without being general intelligence.

11. Challenges, Ethics, and Governance

AI systems can reproduce or amplify bias in historical data. A model trained on unequal or incomplete examples may provide systematically different outcomes for different groups. Fairness cannot be assessed without identifying the relevant populations, decision context, and harms.

Privacy is another major concern. Data collection should have a legitimate purpose, appropriate notice, access controls, retention limits, and protection against unauthorized disclosure. Sensitive data require stronger safeguards.

Other risks include inaccurate generated content, adversarial manipulation, insecure model interfaces, intellectual-property disputes, excessive automation, and unclear responsibility when an AI-assisted decision causes harm. Explainability is especially important where people may be denied opportunities, services, liberty, or safety.

Responsible practice includes:

• defining the purpose and limits of the system;
• using representative, legally obtained, and well-documented data;
• testing performance across relevant conditions and groups;
• measuring both benefits and errors;
• keeping humans meaningfully involved in high-impact decisions;
• documenting model versions, assumptions, limitations, and evaluation results;
• monitoring performance after deployment;
• providing appeal, correction, and incident-reporting processes;
• applying security, privacy, and access controls;
• complying with applicable laws, standards, and organizational policies.

AI governance is not only a technical activity. It involves management, domain experts, legal professionals, affected communities, and users. The appropriate level of oversight depends on the application’s risk, autonomy, scale, and impact.

12. Contemporary Direction and Professional Practice

Modern AI practice increasingly combines machine learning with retrieval systems, tools, knowledge bases, simulation, and human review. Generative models can produce useful drafts, code, images, and summaries, but generated output must be verified. A fluent answer is not automatically a correct answer.

Professionals should learn to ask five practical questions:

1. What decision or task is the system intended to support?
2. What evidence demonstrates that it works in the intended environment?
3. What kinds of error are possible, and who bears the consequences?
4. What data, permissions, infrastructure, and skills are required?
5. What happens when the system is uncertain, unavailable, manipulated, or wrong?

The most effective AI projects usually begin with a well-defined problem and a feasible data strategy. They establish a baseline using a simple method before adopting a more complex model. They also treat deployment, monitoring, documentation, and change management as part of the system—not as activities to be added after development.

Chapter Summary

Artificial intelligence is the field of designing computational agents that perceive information, represent knowledge, learn or reason, and act toward goals. Narrow AI performs bounded tasks and is widely deployed, while AGI and superintelligence remain hypothetical concepts. AI developed through contributions from logic, statistics, neuroscience, psychology, and computing, progressing from symbolic reasoning and expert systems to machine learning, deep learning, and foundation models.

Intelligent agents operate through perception, representation, decision, and action. Rationality means selecting actions expected to perform well according to a defined measure, given available information and resources. PEAS provides a practical way to describe an agent’s task, while environment properties such as observability, uncertainty, dynamics, and continuity influence system design.

Knowledge representation enables AI systems to store and manipulate facts, relationships, rules, probabilities, or learned numerical patterns. Search and planning support problem solving, and reasoning may be deductive, inductive, abductive, probabilistic, causal, or constraint-based. AI applications range from language and vision to healthcare, manufacturing, finance, education, transport, and cybersecurity.

Successful AI requires more than a capable model. Data quality, evaluation, security, privacy, fairness, human oversight, monitoring, and accountability are essential. The central professional lesson is to match the method to the task, assess evidence in the real environment, and manage the consequences of error.

Key Terms

Artificial intelligence: The design of computational systems that perform tasks associated with intelligence.

Agent: An entity that perceives an environment and acts upon it.

Artificial general intelligence: A hypothetical system with broad, transferable intelligence across domains.

Artificial superintelligence: A hypothetical intelligence that exceeds human capability broadly.

Deep learning: Machine learning using multi-layer neural networks.

Environment: The external world or context in which an agent operates.

Foundation model: A broadly trained model adapted to multiple downstream tasks.

Heuristic: A rule or estimate used to guide problem solving or search.

Intelligent agent: A goal-directed agent that selects actions using perception, reasoning, learning, or planning.

Knowledge representation: The encoding of information so that a computer can use it for inference or action.

Machine learning: Methods that improve performance by learning patterns from data or experience.

Narrow AI: AI designed for a limited task or domain.

Ontology: A formal description of concepts and relationships in a domain.

PEAS: Performance measure, environment, actuators, and sensors.

Percept: An input or observation received by an agent.

Rational agent: An agent that selects actions expected to maximize its performance measure.

Search: Systematic exploration of possible states or actions to find a solution.

Symbolic AI: AI based primarily on explicit symbols, rules, and logical manipulation.

Utility: A numerical or ordinal representation of the desirability of an outcome.

Review Questions

Short Answer Questions

1. Define artificial intelligence in terms of agents, environments, and goals.
2. What is the principal difference between narrow AI and artificial general intelligence?
3. Explain the four components of the PEAS framework.
4. Distinguish a fully observable environment from a partially observable environment.
5. What is the purpose of a heuristic in informed search?

Long Answer Questions

1. Trace the major historical developments of AI from formal logic and early neural models to contemporary foundation models.
2. Explain rational agents and compare simple reflex, model-based, goal-based, utility-based, and learning agents.
3. Discuss knowledge representation approaches and the trade-offs between symbolic and statistical representations.
4. Explain how environment properties influence the design of an AI system, using an autonomous vehicle as an example.
5. Discuss the technical, ethical, and governance requirements for deploying AI in a high-impact application.

Analytical / Application Questions

1. Design a PEAS description for an AI system that recommends courses to diploma students. Identify at least two possible harms and propose safeguards.
2. A warehouse robot operates among moving workers, has incomplete sensor information, and must minimize time and battery use. Classify its environment properties and recommend an appropriate agent architecture.
3. A company reports that its image-inspection model is highly accurate in testing but performs poorly after a new product package is introduced. Analyze possible causes and propose a monitoring and corrective-action plan.

Multiple Choice Questions with Answers

Question 1

Which statement best describes narrow AI?

A. It performs a defined task or limited group of tasks.

B. It has human-level ability in every intellectual domain.

C. It necessarily exceeds human intelligence in all domains.

D. It operates without data, rules, or objectives.

Correct Answer: A. It performs a defined task or limited group of tasks.

Question 2

In the PEAS framework, what does the letter S represent?

A. Strategies

B. Sensors

C. States

D. Search

Correct Answer: B. Sensors

Question 3

What makes an agent rational?

A. It always obtains the best possible outcome.

B. It selects actions expected to maximize its performance measure given available information.

C. It imitates human behavior in every situation.

D. It never makes an incorrect prediction.

Correct Answer: B. It selects actions expected to maximize its performance measure given available information.

Question 4

Which environment property means that some relevant information is hidden or noisy?

A. Fully observable

B. Episodic

C. Partially observable

D. Deterministic

Correct Answer: C. Partially observable

Question 5

In the expression f(n) = g(n) + h(n), what does g(n) represent?

A. The estimated remaining cost

B. The cost already incurred

C. The probability of failure

D. The number of available sensors

Correct Answer: B. The cost already incurred

Question 6

Which approach primarily represents knowledge using explicit facts and rules?

A. Symbolic AI

B. Random sampling

C. Image compression

D. Hardware virtualization

Correct Answer: A. Symbolic AI

Question 7

Which learning setting uses labelled examples containing inputs and target outputs?

A. Unsupervised learning

B. Reinforcement learning

C. Supervised learning

D. Evolutionary search only

Correct Answer: C. Supervised learning

Question 8

Why is a false positive important in an automated visual inspection system?

A. It labels a defective item as defective.

B. It labels a good item as defective or sends it for unnecessary review.

C. It guarantees that the model is biased.

D. It prevents the system from receiving sensor input.

Correct Answer: B. It labels a good item as defective or sends it for unnecessary review.

Question 9

Which statement about artificial general intelligence is accurate?

A. It is identical to every current recommendation system.

B. It refers to hypothetical broad, transferable intelligence.

C. It is limited to one classification task.

D. It is another name for a database.

Correct Answer: B. It refers to hypothetical broad, transferable intelligence.

Question 10

Which practice is most appropriate for a high-impact AI deployment?

A. Remove human review to maximize automation.

B. Evaluate only average accuracy before deployment.

C. Use monitoring, documentation, safeguards, and meaningful human oversight.

D. Ignore model behavior after the initial release.

Correct Answer: C. Use monitoring, documentation, safeguards, and meaningful human oversight.