Computer Science › Section L · Artificial intelligence › Topic L.1

Intelligence and machines

Artificial intelligence asks whether machines can perceive, reason, learn and act sensibly on their own, and what it would even mean to say they do. This topic sets out the vocabulary of agents and their levels of behaviour, the two reasons people study AI, and the classic test of machine intelligence with its famous loophole.

What artificial intelligence aims at

Artificial intelligence (AI) is the branch of computer science that builds machines able to carry out tasks that would call for intelligence if a person did them: recognising a face, understanding a spoken request, planning a route or learning from experience. Many come effortlessly to people, which is partly why they proved so hard to program. The field borrows from psychology, linguistics, mathematics, philosophy and engineering. Its name comes from a proposal that John McCarthy and colleagues wrote in 1955 for a summer research workshop at Dartmouth College in 1956, often treated as the founding event of the field.

Agents, sensors and actuators

Much AI research is framed in terms of agents. An agent is anything that takes in data about its surroundings and responds to it. It senses through sensors, such as cameras, microphones, touch pads, thermometers or GPS receivers, and it acts through actuators, such as motors, wheels, grippers, loudspeakers or a screen. An agent need not have a body: a spam filter senses incoming email and acts by choosing a folder. An agent behaves rationally when its actions are sensible responses to what it has sensed, given what it is trying to achieve.

Levels of response

Agents can be compared by how they decide what to do. The simplest give a reflex response: a fixed reaction to the current input, like an automatic door that opens whenever its motion sensor fires. A knowledge-based agent keeps information about its world and uses it, so the same input can lead to different actions in different situations. A goal-based agent looks ahead, forming a plan or choosing the best of several options to reach a goal, as a chess program does. A learning agent improves its responses with experience. Real systems often combine several of these levels.

Two kinds of knowledge

Learning can add two kinds of knowledge. Procedural knowledge is knowing how: how to balance, how to grip a cup, how to play a strong opening. It is usually learned by trial and error, with good actions rewarded and poor ones penalised. Declarative knowledge is knowing what: facts such as 'the kitchen door is locked' or 'the meeting moved to Tuesday'. Learning it means adding or revising facts in the agent's store of knowledge. Keeping those facts current as the world changes is a hard problem in its own right, taken up in 'Additional areas: knowledge and learning'.

Worked example: classifying four agents

Classify each response as reflex, knowledge-based, goal-based or learning. 1. A smoke alarm sounds whenever smoke reaches its sensor. Its output depends only on the current input, so it is a reflex. 2. A home heating controller keeps a record of when each room is usually occupied and warms rooms before people arrive. It acts on stored knowledge of its world, so it is knowledge-based. 3. A delivery drone compares possible routes around no-fly zones and picks the shortest permitted one. It forms a plan to reach a goal, so it is goal-based. 4. A draughts (checkers) program wins more often after playing thousands of games against itself. Its behaviour improves with experience, so it is a learning agent.

Engineering and science: two research tracks

AI is pursued for two reasons. On the engineering track the aim is a system that performs well, whether or not it works the way a mind does: performance-oriented research. On the scientific track the aim is to understand intelligence, especially human intelligence, so a program is a model to be tested against how people behave: simulation-oriented research. A voice-controlled light switch only has to tell 'on', 'off' and 'dim' apart, so an engineer can match each sound to a stored pattern and ignore the rest of English. A scientist modelling how toddlers learn words needs the process itself to resemble a child's.

The Turing test

In his 1950 paper 'Computing Machinery and Intelligence', Alan Turing replaced the vague question 'Can machines think?' with a game. An interrogator exchanges typed messages with two hidden players, one human and one machine, and tries to tell which is which. If interrogators cannot reliably tell, the machine has, by Turing's standard, behaved intelligently. Turing predicted that in about fifty years machines would play so well that an average interrogator would have no better than a 70% chance of making the right identification after five minutes of questioning. The test judges behaviour only, which is both its appeal and its main weakness.

ELIZA and the illusion of understanding

In the mid-1960s Joseph Weizenbaum at MIT wrote ELIZA, a program that held typed conversations. Its best-known script, DOCTOR, played a psychotherapist who turns the patient's statements back into questions. ELIZA understood nothing: it searched each sentence for keywords, applied a matching rule or fell back on a stock reply. Yet many users felt understood and confided in it, a reaction now called the ELIZA effect. Weizenbaum became a lifelong critic of handing human judgement to machines. For the Turing test, the lesson is that seeming intelligent can be cheap.

Worked example: running ELIZA-style rules

Rules, tried in order: R1, 'I am X' becomes 'How long have you been X?'; R2, a sentence mentioning mother, father or family gets 'Tell me more about your family.'; otherwise reply 'Please go on.' Before reusing X, swap pronouns: my becomes your, me becomes you. Input 'I am worried about my exam'. 1. R1 matches with X = 'worried about my exam'. 2. Swapping pronouns gives 'worried about your exam'. 3. Reply: 'How long have you been worried about your exam?' Input 'My father never listens to me'. 4. R1 fails; R2 matches 'father', so the reply is 'Tell me more about your family.' Input 'The bus was late'. 5. No rule matches: 'Please go on.' The replies sound attentive, yet no step used meaning.

Strong AI, weak AI and general AI

Weak AI is the claim that machines can be programmed to behave intelligently, to act as if they think; few dispute it today. Strong AI is the claim that a suitably programmed machine would really have a mind, with understanding and perhaps consciousness. The philosopher John Searle's 1980 'Chinese room' argument holds that following symbol-shuffling rules, however well, is not understanding; critics reply that brains, too, are built from parts that understand nothing on their own. A separate distinction is narrow AI, built for one kind of task, against artificial general intelligence, as flexible as a person across almost any task: widely regarded as not yet achieved, though how close current systems come is debated.

Common mistakes to avoid

First, equating AI with robots: most AI is software, such as spam filters and translation tools. Second, assuming a reflex cannot be rational: for a smoke alarm it is exactly the right response. Third, treating a pass in the Turing test as proof of understanding, when ELIZA shows how little machinery can create that impression. Fourth, confusing weak AI, intelligent behaviour, with strong AI, a real mind. Fifth, assuming performance-oriented systems must copy human thinking; engineers often use shortcuts that work only in a narrow setting.

Key takeaways

AI builds machines for tasks that need intelligence in people. An agent senses through sensors and acts through actuators; its responses range from reflex to knowledge-based, goal-based and learning. Procedural knowledge is knowing how, declarative knowledge knowing what. Performance-oriented research wants systems that work; simulation-oriented research wants to understand minds. Turing's 1950 test asks whether interrogators can tell a machine from a person, and ELIZA showed how readily people see understanding in simple rules. Weak AI concerns behaviour, strong AI a genuine mind.

Flashcards

Flashcard

Artificial intelligence

Building machines that do tasks which would need intelligence if a person did them

Named in the 1955 proposal for the 1956 Dartmouth workshop.

Test your understanding

Question 1 of 5

A garden sprinkler switches on whenever its soil-moisture reading falls below a fixed value. What level of response is this?

Question 2 of 5

Which description matches the Turing test?

Question 3 of 5

Using the ELIZA-style rules from the worked example (R1 tried before R2), what is the reply to 'I am angry with my mother'?

Question 4 of 5

Which statement expresses the claim of strong AI?

Question 5 of 5

Researchers write a program that forgets items from shopping lists and compare its errors with those of volunteers. Which approach is this?

← Previous topicNext topic →