Syllabus topics: Definition and scope of AI; history and evolution of AI; the Turing Test; applications of AI in the real world. Types of AI — Weak AI vs Strong AI, Narrow AI vs General AI. Intelligent agents — structure of agents, rationality, agent types. Environments — deterministic vs stochastic, static vs dynamic, discrete vs continuous. PEAS representation (performance measure, environment, actuators, sensors).
DEFINITION
The definition and scope of AI is the first thing the syllabus asks for, and it is genuinely contested — which is why the exam answer needs a definition and a boundary.
THE BIG IDEA
AI is the study of building systems that act rationally — that do the right thing given what they know.
The word "intelligent" invites arguments that the field has learned to avoid. Russell and Norvig's four-way split is the standard way to organise them, and it is a standard exam question:
| Human-centred | Rationality-centred | |
|---|---|---|
| Thought | Systems that think like humans — cognitive modelling | Systems that think rationally — logic, the "laws of thought" |
| Behaviour | Systems that act like humans — the Turing Test | Systems that act rationally — the rational agent |
The bottom-right cell is the one the field actually pursues, and it is the definition to give. Reasons worth stating:
Rationality is more general than logic. Sometimes the right action cannot be proved correct, and you must act anyway.
It does not require consciousness, which nobody can test for.
IN DEPTH
| Year | Event |
|---|---|
| 1943 | McCulloch and Pitts model an artificial neuron |
| 1950 | Turing's "Computing Machinery and Intelligence" — the Imitation Game |
| 1956 | The Dartmouth Conference. John McCarthy coins "Artificial Intelligence" |
| 1957–69 | Early optimism: the Logic Theorist, GPS, perceptrons, ELIZA |
| 1969 | Minsky and Papert's Perceptrons — kills neural network funding |
| 1970s | The first AI winter — promises unmet, funding cut |
| 1980s | Expert systems — MYCIN, XCON. AI's first commercial success |
| late 80s | The second AI winter — expert systems proved brittle |
| 1997 | Deep Blue beats Kasparov |
| 1990s–2000s | The statistical turn: machine learning, probability, data |
| 2012 | AlexNet — deep learning wins ImageNet decisively |
| 2016 | AlphaGo beats Lee Sedol |
| 2017– | Transformers; large language models |
Two things to say about the winters, because they are the interesting part: both followed over-promising, and both ended when the field found something that actually worked on real problems — expert systems, then statistical learning. The pattern is worth a mark.
FORMULA
A human interrogator converses by text with a human and a machine. If the interrogator cannot reliably tell which is which, the machine passes.
Turing's move was to replace an unanswerable question ("can machines think?") with an operational one.
| Requires the machine to have | Because |
|---|---|
| Natural language processing | To communicate at all |
| Knowledge representation | To store what it knows |
| Automated reasoning | To answer and draw conclusions |
| Machine learning | To adapt and generalise |
| (Total Turing Test) computer vision and robotics | To perceive and manipulate objects |
It tests imitation, not intelligence. A system can pass by being evasive and making deliberate mistakes; ELIZA (1966) fooled people with pattern substitution and no understanding at all.
Searle's Chinese Room. A person following rule-books to manipulate Chinese symbols could pass while understanding nothing — so behaviour does not establish understanding.
It is anthropocentric. Aircraft do not flap. Requiring machines to be indistinguishable from humans is not a useful engineering target.
It is not a research programme. Nobody builds systems by aiming at it.
Its value is historical and philosophical, and saying that is the mature answer.
FORMULA
The first axis is Narrow AI against General AI — how wide the capability is:
| Narrow AI (Weak) | General AI (AGI) | Super AI | |
|---|---|---|---|
| Scope | One task | Any intellectual task a human can do | Beyond human, across the board |
| Exists? | Yes — everything today | No | No |
| Examples | Chess engines, spam filters, translation, self-driving | — | — |
That axis is about capability breadth. The other is about claims of mind:
| Weak AI (hypothesis) | Strong AI (hypothesis) | |
|---|---|---|
| Claims | The machine acts as if it thinks. A useful tool | The machine genuinely thinks and has a mind |
| Is it a claim about | Behaviour | Consciousness |
| Testable? | Yes | Not agreed to be |
| Searle's target | — | Strong AI — the Chinese Room argues against exactly this |
In Searle's sense: weak = simulating thought, strong = really thinking. A philosophical distinction.
In common usage: weak = narrow, strong = general. A capability distinction.
Say which sense you are using. A good answer gives both tables and notes that the terms overlap confusingly — that observation is itself worth a mark.
| Domain | Application | Technique |
|---|---|---|
| Healthcare | Diagnosis from images; drug discovery; triage expert systems | Vision, expert systems |
| Finance | Fraud detection, algorithmic trading, credit scoring | Classification |
| Transport | Self-driving, route planning, traffic control | Search, vision, RL |
| Language | Translation, chatbots, summarisation | NLP |
| Games | Chess, Go, poker | Search + evaluation, RL |
| Robotics | Manufacturing, surgery, warehouses | Planning, control |
| Agriculture | Yield prediction, disease detection, irrigation | Vision, regression |
| Education | Adaptive tutoring, automated grading | Modelling, NLP |
KEY INSIGHT
Route planning, game playing and puzzle solving are search (Units 2–3). Diagnosis and triage are expert systems (Unit 5). Those are the applications to name when the question is about this course rather than AI in general.
Two questions, in order: what is an agent, and what is the structure of agents — the internal organisation that turns percepts into actions.
THE BIG IDEA
An agent is anything that perceives its environment through sensors and acts on it through actuators.
┌──────────────────────────────────┐
│ AGENT │
│ ┌────────────────────────┐ │
─────┼──►│ agent function │ │
percepts │ f : percept* → action │ │
│ └───────────┬────────────┘ │
│ ▼ │
│ actions ───────────────┼─────►
└──────────────────────────────────┘
ENVIRONMENT
| Term | Means |
|---|---|
| Percept | One input at one instant |
| Percept sequence | Everything perceived so far |
| Agent function | The mapping from percept sequence to action — abstract |
| Agent program | The implementation of that function — concrete, finite |
| Actuator | What the agent acts with |
⚠️ The agent function is the specification; the agent program is the code. The function is a mathematical object, potentially an infinite table; the program is what you actually write. That distinction is a two-mark question.
FORMULA
A rational agent selects, for each percept sequence, the action expected to maximise its performance measure, given the evidence provided by the percept sequence and whatever built-in knowledge it has.
Four things it depends on, and the definition names all four:
| Means | |
|---|---|
| Omniscient | Knows the actual outcome. Impossible |
| Rational | Maximises expected performance, given what is knowable |
Crossing a road after looking both ways is rational even if a cargo door falls from a passing aeroplane and kills you. Rationality is about the decision, not the outcome — and that example is worth using, because it makes the point instantly.
Two related ideas:
Information gathering — doing something to improve your percepts (looking before crossing) is rational.
Learning — a rational agent should improve from experience, rather than relying only on built-in knowledge. An agent that cannot is said to lack autonomy.
FORMULA
| Type | Decides from | Keeps state? | Handles |
|---|---|---|---|
| Simple reflex | The current percept only — condition–action rules | No | Fully observable environments only |
| Model-based reflex | Percept + internal state of how the world evolves | Yes | Partial observability |
| Goal-based | State + a goal — needs search or planning | Yes | Choosing between actions toward a goal |
| Utility-based | State + a utility function over outcomes | Yes | Conflicting goals, and degrees of success |
| Learning | Any of the above, plus improvement from experience | Yes | Unknown environments |
KEY INSIGHT
Simple reflex fails the moment the right action depends on something not currently visible. A vacuum agent that cannot see the other square loops for ever.
Model-based fixes that by remembering — an internal model of how the world evolves and how actions affect it.
Goal-based is needed because knowing the state does not say what to do. This is where search enters, and where Units 2 and 3 live.
Utility-based is needed because goals are binary — reached or not — and real problems have trade-offs: fastest route versus safest versus cheapest. A utility function makes them comparable.
Learning is needed because you cannot programme every situation in advance.
FORMULA
| Component | Does |
|---|---|
| Learning element | Makes improvements |
| Performance element | Selects actions — this is "the agent" in the earlier sense |
| Critic | Reports how well the agent is doing against a fixed standard |
| Problem generator | Suggests exploratory actions — deliberately suboptimal, to learn something |
The problem generator is the one people forget, and it is the interesting one: without deliberate exploration the agent only ever refines what it already does. That is the exploration–exploitation trade-off of Machine Learning §1.3, arriving from the other direction.
FORMULA
| Letter | Stands for | Answers |
|---|---|---|
| P | Performance measure | What counts as doing well? |
| E | Environment | What is it operating in? |
| A | Actuators | What can it do? |
| S | Sensors | What can it perceive? |
FORMULA
An automated taxi driver:
| P | Safety, legality, speed, comfort, fuel economy, profit |
| E | Roads, other traffic, pedestrians, weather, signs, passengers |
| A | Steering, accelerator, brake, indicators, horn, display |
| S | Cameras, LIDAR, GPS, speedometer, accelerometer, engine sensors |
A medical diagnosis expert system:
| P | Correct diagnosis, patient health, cost, time to diagnosis |
| E | Patient, hospital staff, test facilities |
| A | Questions, test requests, diagnoses, treatment recommendations |
| S | Typed symptoms, test results, patient history |
Design it for what you want in the environment, not for how you think the agent should behave.
The standard example: reward a vacuum agent for the amount of dirt collected, and a rational agent will dump the dirt out and collect it again. It is maximising exactly what you asked for. Reward a clean floor instead.
This is the alignment problem in miniature, and it connects to Unit 5's AI ethics. Worth stating.
FORMULA
| Dimension | Example of the harder case | ||
|---|---|---|---|
| Observability | Fully — sensors give the complete state | Partially — some state hidden | Poker: you cannot see other hands |
| Determinism | Deterministic — next state fixed by current state and action | Stochastic — probabilistic | Driving: a tyre may burst |
| Episodic vs sequential | Episodic — each decision independent | Sequential — actions have long-term consequences | Chess: an early move decides the endgame |
| Static vs dynamic | Static — unchanged while you deliberate | Dynamic — changes while you think | Driving: traffic does not wait |
| Discrete vs continuous | Discrete — finite states and actions | Continuous | Steering angle, speed |
| Single vs multi-agent | Single | Multi — competitive or cooperative | Chess (competitive), driving (both) |
| Known vs unknown | The rules are known | The rules must be learned | A new game |
Known ≠ observable. Known is about whether you understand the rules; observable is about whether the sensors show the state. A new board game with everything visible is fully observable but unknown; solitaire played by an expert is known but partially observable.
Deterministic ≠ certain to succeed. Stochastic means genuinely probabilistic outcomes. An environment that is deterministic but partially observable can look stochastic from inside — and Russell and Norvig call that nondeterministic rather than stochastic.
FORMULA
| Environment | Observable | Deterministic | Episodic | Static | Discrete | Agents |
|---|---|---|---|---|---|---|
| Crossword | Fully | Deterministic | Sequential | Static | Discrete | Single |
| Chess with a clock | Fully | Deterministic | Sequential | Semi | Discrete | Multi |
| Poker | Partially | Stochastic | Sequential | Static | Discrete | Multi |
| Taxi driving | Partially | Stochastic | Sequential | Dynamic | Continuous | Multi |
| Medical diagnosis | Partially | Stochastic | Sequential | Dynamic | Continuous | Single |
| Image classification | Fully | Deterministic | Episodic | Semi | Continuous | Single |
Taxi driving is the hardest case on every dimension, which is why it is the standard example — and why it is still not solved.
KEY INSIGHT
It tells you which algorithms are even applicable.
Fully observable, deterministic, discrete, static → the search of Units 2 and 3 works directly.
Partially observable → you need belief states or a model.
Say that in the exam. The table alone is description; the consequence is the answer.
PROBLEM 1
What is the Turing Test? Explain its structure, what it requires, and the main criticisms. (10 marks)
Solution.
The setup: a human interrogator converses by text with two respondents, one human and one machine. If the interrogator cannot reliably tell which is which, the machine passes. Turing proposed it in 1950 to replace the unanswerable "can machines think?" with an operational question.
What passing requires: natural language processing, knowledge representation, automated reasoning and machine learning — plus, in the Total Turing Test, computer vision and robotics.
The criticisms, and give at least three:
It tests imitation, not intelligence. ELIZA (1966) fooled people with pattern substitution and no understanding.
Searle's Chinese Room: a person following rule-books to manipulate Chinese symbols could pass while understanding nothing, so behaviour does not establish understanding.
It is anthropocentric. Aircraft do not flap; requiring indistinguishability from humans is not a useful engineering target.
It is not a research programme. No serious system is built by aiming at it.
Conclude maturely: its value is historical and philosophical. The field pursues acting rationally instead, because that is measurable, more general than logic, and does not require consciousness.
PROBLEM 2
Explain the five types of intelligent agent, and why each was needed. (10 marks)
Solution.
Give the table — simple reflex, model-based reflex, goal-based, utility-based, learning — with what each decides from and whether it keeps internal state.
Then the progression, which is the part that earns the marks:
Simple reflex acts on the current percept alone via condition–action rules. It fails the moment the right action depends on something not currently visible — a vacuum agent that cannot see the other square loops for ever.
Model-based fixes that with an internal state plus a model of how the world evolves and how actions affect it. It handles partial observability.
Goal-based is needed because knowing the state does not say what to do. This is where search enters, and it is Units 2 and 3.
Utility-based is needed because goals are binary while real problems have trade-offs — fastest against safest against cheapest. A utility function makes outcomes comparable.
Learning is needed because you cannot programme every situation in advance.
Finish with the learning agent's four components — learning element, performance element, critic, problem generator — and note that the problem generator deliberately suggests suboptimal exploratory actions, because without exploration the agent only refines what it already does.
PROBLEM 3
What is PEAS? Give the PEAS description of an automated taxi. Why is the performance measure the hard part? (10 marks)
Solution.
PEAS specifies an agent's task environment: Performance measure, Environment, Actuators, Sensors.
Give the taxi table in full — P: safety, legality, speed, comfort, fuel economy, profit; E: roads, traffic, pedestrians, weather, signs, passengers; A: steering, accelerator, brake, indicators, horn, display; S: cameras, LIDAR, GPS, speedometer, engine sensors.
Then classify the environment and note it is the hardest case on every dimension: partially observable, stochastic, sequential, dynamic, continuous, multi-agent.
Why the performance measure is hard: design it for what you want in the environment, not for how you think the agent should behave. Reward a vacuum agent for dirt collected and a rational agent will dump the dirt out and collect it again — it is maximising exactly what you asked for. Reward a clean floor instead.
That is the alignment problem in miniature, and it links to Unit 5's ethics material. Note too that the taxi's own measures conflict — speed against safety against comfort — which is precisely why a taxi needs a utility-based agent rather than a goal-based one.
Two marks
Five marks
Ten marks
COMMON ERRORS
Confusing weak/strong with narrow/general. They are different axes, and "weak AI" has two established meanings. Say which you mean.
Saying a rational agent always succeeds. It maximises expected performance. Looking both ways is rational even if you are still hit.
Giving four agent types. There are five; the learning agent is usually the one dropped.
PEAS with a vague performance measure. "Drive well" earns nothing; "safety, legality, speed, comfort, fuel economy, profit" earns the mark.
Confusing "known" with "observable". Known is about the rules; observable is about the sensors.
Listing environment properties with no consequence. The table is description; which algorithms become applicable is the answer.
Dating the Turing Test to the Dartmouth Conference. Turing 1950; Dartmouth 1956.