Introduction

These notes were written for COMP5001. They follow the Week 1 lecture and Chapters 1 and 2 of Russell and Norvig’s Artificial Intelligence: A Modern Approach (4th ed.). The definition of an agent and the list of topics come from CS50’s Introduction to Artificial Intelligence with Python by Brian Yu and David J. Malan (Harvard University), licensed under CC BY-NC-SA 4.0. The notes are shared under the same licence.

1. The module
2. What is AI?
3. Four goals for AI
4. Agents
5. Describing the task: PEAS and environments
6. Five kinds of agent
7. The eight topics
8. Symbolic and statistical AI
9. Where the ideas came from
10. A short history of AI
11. Hype and winters
12. What can AI do today?
13. Benefits and risks
14. The Week 1 puzzles, and before Week 2
15. Check yourself, key terms and reading

The ten weeks

Week 1 is this introduction. Weeks 2 to 9 each take one topic (the table in Section 7). Week 10 is AI in practice and revision, with a portfolio clinic. Every week has three parts:

The portfolio

The module has one assessment: a portfolio worth 100% of the mark. It is set around Kingsgate Couriers, a made-up parcel company in Hampshire whose managing director wants to know where AI could help.

Deadline: 3.30pm, Wednesday 6 January 2027. The practicals prepare you for Part 1 and the seminars for Part 2. For Part 3, keep a few lines each week on what you learned and what you found hard. The assessment brief on the module page has the details.

Why? The portfolio tests whether you can build a small AI system, explain how it works, and judge where AI fits in an organisation. An AI tool would be doing that work for you. Before you start, switch off AI code completion in your editor (for example GitHub Copilot in VS Code): some editors switch it on by default.

Reading and help

• Russell, S. and Norvig, P. (2020) Artificial Intelligence: A Modern Approach. 4th edn. Pearson.
• Poole, D.L. and Mackworth, A.K. (2023) Artificial Intelligence: Foundations of Computational Agents. 3rd edn. Cambridge University Press. Free online at artint.info.

The end of each week’s notes lists the sections to read. Office hours: Dr Sakib Anwar, UWBS 005, Wednesdays 4 to 5pm, in 30-minute slots booked at calendly.com/sakib-anwar. Email: sakib.anwar@winchester.ac.uk.

Recognising a friend’s face, understanding a spoken question, finding the fastest way across town, playing chess and deciding whether an email is a scam all need intelligence when a person does them. A computer that does them is doing AI.

The line also moves over time. Finding a route and playing chess were headline AI once; now most people call them ordinary software. This is sometimes called the AI effect: once a computer can do something reliably, people stop calling it AI. The seminar’s “AI or not AI?” vote explores where the line is today.

Act like people: the Turing test

In 1950 the British mathematician Alan Turing asked “Can machines think?” That question is hard to answer, so he suggested a test instead, which he called the imitation game. It is now called the Turing test.

To pass, a machine needs to understand and write language, store what it knows, reason with that knowledge to answer questions, and learn from the conversation. Those became four of the main areas of AI. Turing predicted that by 2000 a computer would fool an average interrogator about 30% of the time after five minutes of questions.

The test has weaknesses. It rewards imitation, not intelligence: in 1966 a simple program called ELIZA (Section 10) fooled some people with a few tricks. The result also depends on who asks the questions, so two runs of the test can disagree. In a 2025 study, people judged a large language model to be the human 73% of the time in five-minute chats. Few researchers take that as proof that it thinks.

Think like people: cognitive science

To build a program that thinks the way people do, we first need to know how people think. That comes from psychology experiments and from studying the brain, and it is the subject of cognitive science. It is now a separate field from AI, but ideas pass both ways: neural networks (Week 8) were inspired by brain cells.

Think logically: the laws of thought

Over 2,000 years ago the Greek philosopher Aristotle wrote down patterns of correct reasoning, such as: “Socrates is a man. All men are mortal. So Socrates is mortal.” Each line is a statement: a sentence that is either true or false. If the first two statements are true, the conclusion must be true. Logic like this is the language of symbolic AI (Section 8). Week 3 writes statements like these in symbols and works out what follows from them, and Week 4 uses them to describe facts and actions for a planner.

Logic alone is not enough, for two reasons. Much of what we know is uncertain (“it will probably rain”), which Week 5 handles with probability. And much intelligent behaviour is not reasoning at all: you pull your hand off a hot pan before you have time to think.

Act to get the best result: rational agents

A rational agent does the right thing: the action that is expected to achieve its goals best, given what it knows. It does not have to think like a person, and it does not have to use logic. Even a reflex can be rational, if it gets the best result. Section 4 shows how we score an agent to decide what “best” means.

Perfect rationality is usually out of reach, because there is not enough time to work out the perfect answer. A chess program cannot check every possible game (Week 2). So in practice we look for the best action we can find with the time and computer we have. Week 3’s heuristics, quick guesses that point a search in the right direction, are one way to do this.

4Agents

Russell and Norvig add how an agent does this. It perceives through sensors: cameras, microphones, touch sensors, or simply the data it is sent. It acts through actuators: wheels, motors, a speaker, a screen, or a message it sends. The environment is the world the agent is in.

What the agent senses at one moment is a percept. Everything it has sensed so far is its percept sequence. The agent’s job is to choose the next action, using its percept sequence.

A sat-nav, a robot vacuum, a spam filter and a chess app are all agents. So is a thermostat: it senses the room temperature and switches the heating on or off. But is it intelligent? It follows one fixed rule (“colder than the setting: heating on”). It never compares options, plans ahead or learns. Being an agent is not the same as being intelligent: the AI is in how well the agent chooses its actions.

Worked example: the vacuum world

Russell and Norvig use a tiny world to show these ideas. There are two squares, A and B, and a robot vacuum. Each square is either clean or dirty.

The robot can sense two things: which square it is in, and whether that square is dirty. It cannot see the other square. It can do four things: move Left, move Right, Suck up the dirt, or do nothing.

We write a percept as a pair in square brackets. How to read it: [A, Dirty] means “I am in square A, and square A is dirty”. A table can then list the action to take for each percept:

In Figure 3 the percept is [A, Dirty], so the robot sucks. A complete table would need a row for every possible sequence of percepts ([A, Clean] then [A, Dirty], and so on), so it would never end. Instead we write a short rule that gives the same answers:

Rational, not perfect

Is this a good vacuum? To answer, someone has to decide how to score it: its performance measure. In the vacuum world we could give one point for each clean square at each moment. Or we could also take away a point for every move, to save electricity. With the second score, a vacuum that keeps moving between two clean squares loses points, so the best behaviour changes. People choose the score, and different scores lead to different “best” actions. Week 2 meets this again, when a sat-nav has to decide whether the best route is the fastest, the shortest or the one that uses least fuel.

Rational does not mean all-knowing, and it does not mean always successful. You look both ways, the road is clear, and you cross. Halfway over, a part falls off a passing plane and hits you. Crossing was still the rational choice: nobody could have known. A rational agent makes the best choice with the information it has; it cannot see the future.

So a rational agent should gather information before it acts (looking before crossing), learn from what it senses, and be autonomous: able to act on its own experience, not only on what its designer built in.

Writing the PEAS first stops you building the wrong thing. Compare the taxi with the vacuum world: its performance measure alone has five parts, and some pull against each other (fast against safe).

Six questions about an environment

Some environments are much harder for an agent than others. Six questions show how hard. In each row, the middle column is the easier answer for the agent and the right-hand column is the harder one.

A self-driving taxi is at the harder end of every row. So is most of the real world: partly hidden, uncertain, sequential, always changing, continuous and full of other agents.

1. Simple reflex agent. It chooses using only what it senses now, with if–then rules (“if the square is dirty, then suck”). The vacuum’s three-line rule and the thermostat are simple reflex agents. They are fast and easy to check, but they only work when the current percept tells them everything they need. Our vacuum cannot see the other square, so once both squares are clean it keeps moving from A to B and back for ever, wasting electricity.

2. Model-based agent. It also keeps track of the parts of the world it cannot sense right now: an internal picture, or model, of the world. It updates the model using what it knows about how the world changes and what its own actions do. A model-based vacuum remembers that it cleaned B a minute ago, so it can wait in A instead of wandering. A driver remembers the car that has just gone into the blind spot. Weeks 3 and 4 write such a picture as a list of facts and reason with it using logic, and Week 5 adds probability.

3. Goal-based agent. It knows what it is trying to achieve, and it looks ahead: “what will happen if I do this?” A sat-nav’s goal is your destination, and it looks for actions that get you there. To look ahead, it must know what each of its actions does: in Week 4 that is written as each action’s preconditions and effects. This is search and planning, in Weeks 2 to 4.

4. Utility-based agent. A goal is either reached or not. But there are usually many ways to reach it, and some are better than others: faster, cheaper or safer. A utility-based agent gives each outcome a score, its utility, and picks the action with the best score. When it cannot be sure what will happen, it picks the best score on average. Minimax in Week 2 uses utilities, and so do Weeks 5 and 6.

5. Learning agent. Any of the four can be made into a learning agent. One part of it, the critic, checks how well the agent is doing against the performance measure, and the agent changes its rules to do better. A spam filter that is retrained when people mark its mistakes is a learning agent. Weeks 7 to 9 are about learning.

Part 2 of the portfolio asks you to pick the right method for each of the company’s problems, so you need to know what each topic is for, not only how it works.

The middle weeks sit between the two: Week 5 brings in probability, and the optimization methods of Week 6 are used by both families.

The story of AI itself starts just before 1950. In 1943 Warren McCulloch and Walter Pitts described a simple mathematical model of a brain cell: the first artificial neuron, and the seed of the neural networks in Week 8.

Notice the swing between the two families in Section 8. From the 1950s to the 1980s, most of the excitement was about symbolic AI: logic, search, planning and expert systems. The perceptron was an early exception. Since the 1990s, and especially since 2012, statistical AI, which learns from data, has led. AlphaGo shows that the best results often come from combining the two.

1. Big promises attract money and attention.
2. Results fall short of the promises.
3. Funding is cut and interest fades: an AI winter.
4. Work carries on quietly, until a new success starts the cycle again.

In 1958 Herbert Simon and Allen Newell predicted that within ten years a computer would be world chess champion. It took until 1997, nearly forty years.

The first winter (mid-1970s) came when early programs that solved small puzzles failed on real problems. The number of possibilities grows far too fast as a problem gets bigger, so a method that checks them all soon runs out of time. Remember the Week 1 seminar: would your maze method still work on a maze with a million squares? Week 3 shows how fast this happens: with 10 true-or-false symbols there are 1,024 possible worlds to check, and with 30 about a billion. The Lighthill Report made this criticism in 1973, and funding was cut in the UK and the USA.

The second winter (late 1980s) came after the expert systems boom. Expert systems were expensive to build and to keep up to date, and they failed on cases that their rules did not cover. Many AI companies went out of business.

Today there is huge investment, and big promises again. Nobody knows whether a third winter will come. What you can do is judge each claim on its own.

These answers change fast: several have moved from “no” to “yes” since Russell and Norvig’s 2020 edition. Always check the date of anything you read about what AI can do.

It also brings risks. Russell and Norvig list these:

Getting what you asked for

Some researchers also worry about the long term: machines that become far more capable than people. Russell and Norvig call this the gorilla problem. Humans and gorillas share an ancestor, but today the gorillas’ future depends on what humans decide, not on what gorillas want. Others think such machines are a long way off. Either way, there is a lesson for today’s systems.

In the Greek myth, King Midas asked that everything he touched would turn to gold. He got exactly what he asked for, and then could not eat or drink. This is the King Midas problem: an AI system that does exactly what we asked, rather than what we meant. The robot vacuum in Section 4, tipping out dirt to suck it up again, is a small example.

In the seminar you choose an organisation (a supermarket, a bank, the NHS or a university), find three uses of AI, and give a benefit, a risk and who is affected for each. That is the start of Part 2 of the portfolio.

In every puzzle you had to keep track of possibilities: squares still to try, states already visited, worlds still possible, days still free. When you solved the maze, you probably headed straight for B rather than trying every square: Week 3 turns that idea into greedy best-first search and A*. But looking at the picture only works for small puzzles. A method written down step by step, an algorithm, can be followed by a computer on a maze with a million squares. Weeks 2 to 6 are about such algorithms. The full solutions to the puzzles are in the Week 1 seminar solutions on the module page.

1. Which of the four goals for AI does the Turing test measure? Give one weakness of the test.
2. Write a PEAS description for a robot lawnmower. Then suggest a badly chosen performance measure, and say what could go wrong.
3. Which kind of agent (Section 6) is each of these: a smoke alarm; a sat-nav; a robot vacuum that remembers which rooms it has already cleaned; a spam filter that is retrained every week?
4. Use the six questions in Section 5 to describe the environment of a program that plays poker.
5. A bank wants to decide who gets a loan. Describe a symbolic approach and a statistical approach, and give one advantage of each.
6. Put these in order, earliest first: Deep Blue beats Kasparov; the Dartmouth workshop; AlexNet; the Lighthill Report; ELIZA.
7. What is an AI winter? Give one reason for each of the two winters.
8. Why is unloading any dishwasher harder for AI than beating the world chess champion?
9. A sat-nav company says its new AI “saves drivers 20% of their time”. What would you ask before believing it?

Key terms

Reading and next week

• Russell & Norvig, Artificial Intelligence: A Modern Approach (4th ed.): Chapter 1, sections 1.1 to 1.5 (what AI is, its foundations, its history, the state of the art, and risks and benefits), and Chapter 2, sections 2.1 to 2.4 (agents and environments, rationality, PEAS and environment types, kinds of agent).
• Poole & Mackworth, Artificial Intelligence: Foundations of Computational Agents (3rd ed., free at artint.info): sections 1.1 to 1.3 (what AI is, its history, agents in environments), 1.5 (the agent design space, close to the six questions in Section 5) and 1.7 (social impact).

Next week: search. You will describe a problem as a search problem, and learn depth-first and breadth-first search. One of them is probably the maze method you wrote in the seminar. Then we look at games, where an opponent fights back.