[6.3.1–3] AI, expert systems, ML basics

What is Artificial Intelligence (AI)?

A practical definition for IGCSE

Artificial Intelligence (AI) refers to computer systems that perform tasks that normally require human intelligence. These tasks include recognising patterns, understanding natural language, making decisions under uncertainty, learning from experience, and solving problems using strategies or rules. In everyday life, AI helps route your journey, recommends videos or music, detects spam messages, and supports doctors in identifying illnesses from scans.

At IGCSE level, you should understand the key characteristics that make a system “intelligent”: it uses data and/or rules, it can reason (apply logic or probabilities to reach conclusions), and in many modern systems it can learn (improve performance as more data becomes available). You should also know the basics of two foundational AI approaches covered in the syllabus: expert systems (rule-driven) and machine learning (data-driven).

Main Characteristics of AI systems

Data and/or rules, reasoning, and learning

AI systems vary widely, but most can be described using three overlapping characteristics:

  • Data and/or rules: Some systems rely on explicitly programmed rules (e.g. “IF symptom A AND symptom B THEN suggest test C”). Others rely on data (large collections of examples) from which patterns are learned. Many practical systems use a combination of both.
  • Reasoning: The system applies logical inference or statistical methods to reach a conclusion or choose an action. Reasoning may be deterministic (clear logical steps) or probabilistic (weighing likelihoods).
  • Learning: The system can adjust its internal parameters or rules based on experience, often improving accuracy over time. Learning may be supervised, unsupervised, or reinforcement-based (see the tabs below).
Learning approaches (contrasting scenarios)

The system learns from labelled data where each input has a known correct answer. Example: recognising whether an email is spam or not spam using past emails already labelled by users. The model adjusts until it predicts the correct label reliably on new emails.

Strength: high accuracy when labels are reliable. Limitation: needs a large, representative labelled dataset.

The system is given unlabelled data and asked to find structure or groups. Example: a retailer groups customers with similar purchasing habits to tailor promotions. No “correct” answers are supplied; the algorithm finds patterns.

Strength: reveals hidden relationships. Limitation: “clusters” may not always match meaningful, real-world categories without expert interpretation.

An agent learns by interacting with an environment. It tries actions, receives rewards or penalties, and updates its strategy to maximise long-term reward. Example: a robot learns to navigate a room without collisions.

Strength: good for sequential decision-making. Limitation: can be slow to train and risky without safeguards.

Expert Systems

Rule-based AI that captures human expertise

An expert system is an AI that uses a knowledge base of facts and rules to emulate the decision-making of a human expert in a particular domain (e.g. medical triage, financial checks, technical support). Rather than learning from data, an expert system typically relies on knowledge engineers who interview experts and encode their reasoning as IF-THEN rules.

Core components
ComponentPurpose
Knowledge baseStores facts about the problem domain and rules that link conditions to conclusions or actions.
Inference engineApplies rules to the facts to derive new facts or recommendations. May use forward chaining (data → conclusion) or backward chaining (goal → what rules must be true).
User interfaceCollects user inputs (symptoms, settings, constraints) and presents the system’s conclusions or advice.
Explanation facilityShows how a conclusion was reached (e.g. “because Rule 12 and Rule 27 fired”). Improves trust and transparency.
Knowledge acquisitionProcess and tools for obtaining and updating rules from human experts.
Reasoning strategies (comparative tabs)

The inference engine starts from known facts and repeatedly applies rules whose conditions are met to generate new facts. This continues until a conclusion is reached. Suitable when many inputs are available and we want to see all the consequences.

The system starts from a target goal (e.g. “Is diagnosis X true?”) and works backwards to check which rules would prove it, asking the user for information as needed. Useful when there are many possible goals but limited data initially.

Many practical systems mix both approaches, using data-driven steps to gather easy conclusions and goal-driven steps when testing specific hypotheses. Hybrid reasoning helps control complexity and the number of questions asked.

Advantages: transparent decisions, uses proven expert knowledge, good where data is scarce but rules are clear. Limitations: difficult to keep rules up to date; struggles with noisy, ambiguous, or entirely new situations outside the encoded knowledge.

Machine Learning (ML)

Data-driven AI that learns patterns

Machine learning is a branch of AI where systems improve their performance at a task by learning from data rather than relying solely on hand-written rules. For example, a system can learn to recognise handwritten digits after being trained on thousands of images. The learned model captures patterns (shapes, edges, frequencies) that distinguish one digit from another.

Typical ML workflow (progression tabs)

Gather a representative dataset, clean it (remove errors, handle missing values), and split into training and testing sets. Good data quality is crucial: biased or unbalanced data leads to biased models.

Choose an algorithm (e.g. decision tree or simple neural network), train it on the training data, and tune settings. Evaluate with validation metrics (e.g. accuracy, precision/recall). Avoid overfitting by checking performance on unseen data.

Integrate the model into a real system. Monitor accuracy over time and update the model as behaviour or data patterns change. Watch for edge cases (rare or unusual inputs) and maintain a way to fall back to safe defaults.

Expert systems vs Machine learning: Expert systems are rule-rich and data-light, giving clear explanations but limited adaptability. ML systems are data-rich and rule-light, adapting well but sometimes offering less transparent reasoning. Many real-world applications combine them: rules provide safety and compliance; ML provides flexible pattern recognition.

Ethical and Practical Considerations

Fairness, transparency, safety

  • Bias: If training data reflects past unfairness, the model may repeat it. Use diverse datasets and regular audits.
  • Explainability: Users should understand, at least at a high level, why a decision was made—especially in education, healthcare, or finance.
  • Privacy: AI often needs large datasets. Data should be collected lawfully, minimised, and protected with appropriate security.
  • Safety and reliability: Have safeguards, fallbacks, and human oversight for high-stakes decisions.

Key Terminology

TermDefinition
AISystems that perform tasks requiring human-like intelligence such as perception, reasoning, and decision-making.
Expert systemRule-based AI that uses a knowledge base and inference engine to reach conclusions.
Inference engineSoftware that applies rules to facts to infer new facts or decisions.
Machine learningAI techniques where models learn patterns from data to make predictions or decisions.
Supervised learningLearning from labelled input-output pairs.
Unsupervised learningLearning structure from unlabelled data (e.g. grouping similar items).
Reinforcement learningLearning by trial and error using rewards and penalties to shape behaviour over time.
OverfittingWhen a model learns noise and performs poorly on new, unseen data.

 Key Takeaways

  • AI systems use data and/or rules, apply reasoning, and often include learning to improve performance.
  • Expert systems are rule-based with clear explanations; they suit stable domains with well-understood expertise.
  • Machine learning learns patterns from data and adapts to new examples, but needs careful evaluation to avoid bias and overfitting.
  • Supervised, unsupervised, and reinforcement learning address different problem types—classification, pattern discovery, and sequential decisions respectively.
  • In practice, combining rules with ML often gives robust, safe, and flexible solutions.
  • Ethical design matters: fairness, transparency, privacy, and safety must be considered from the start.