Skip to content
9618Paper 3 · Advanced Theory§18.1

18. Artificial Intelligence

Graphs and graph algorithms (Dijkstra, A*), machine learning categories (supervised, unsupervised, reinforcement), deep learning, back-propagation and regression.

Statometer40RegularNext Paper 380%
Marks a paper5.6 · 8%Rank#7 of 8 · #5 on P3Trend · last 12Oct/Nov 24 · 31: 0 marksOct/Nov 24 · 32: 0 marksOct/Nov 24 · 33: 0 marksMay/Jun 25 · 31: 6 marksMay/Jun 25 · 32: 6 marksMay/Jun 25 · 33: 6 marksOct/Nov 25 · 31: 6 marksOct/Nov 25 · 32: 6 marksOct/Nov 25 · 33: 5 marksMay/Jun 26 · 31: 9 marksMay/Jun 26 · 32: 9 marksMay/Jun 26 · 33: 8 marks
8 in 10 chance in the next paper

Everything for this topic — study hub

A2 Level · 9618 · Paper 3

Statometer — what 32 real papers say about this topic and each of its 2 syllabus bullets

Regular · #7 of 8 in A2 Level · recomputed with every new session

40REGULAR
Regular#7 of 8 in A2 Level#5 on Paper 3 Steady

Comes up most sessions for a few marks — know the definitions and one worked example.

Next Paper 3
80%
8 in 10 chance it is set
Marks a paper
5.6 / 75
8% of Paper 3 · fair share 17%
Appeared in
26 / 32
Paper 3 sittings 20212026
Last set
May/Jun 2026
9618/33 · Q9 · 8 marks · 3-series streak
Marks in each of the last 12 Paper 3 sittingsOct/Nov 24May/Jun 26
Oct/Nov 24 · 31: 0 marksOct/Nov 24 · 32: 0 marksOct/Nov 24 · 33: 0 marksMay/Jun 25 · 31: 6 marksMay/Jun 25 · 32: 6 marksMay/Jun 25 · 33: 6 marksOct/Nov 25 · 31: 6 marksOct/Nov 25 · 32: 6 marksOct/Nov 25 · 33: 5 marksMay/Jun 26 · 31: 9 marksMay/Jun 26 · 32: 9 marksMay/Jun 26 · 33: 8 marks

What the papers say

  • Set in 26 of 32 Paper 3 sittings — about 8 papers in 10.
  • Worth about 5.6 marks a paper (8% of Paper 3, well under its fair share of 17%).
  • Last set May/Jun 2026 · 9618/33 · Q9 for 8 marks — in the most recent series.
  • Steady at around 5.3 marks a paper year on year.
  • Lives on “Explain” and “Show” — 69% of its questions: full sentences with a reason, not one-word answers.
  • 41% of its questions involve a diagram, table or figure — practise with pen and paper.
  • 41% of its questions are set out as code, pseudocode or a table to complete.
  • Its biggest question so far: 10 marks (May/Jun 2023 · 9618/32 · Q11).
  • Inside the topic, §18.1 Graphs for AI carries the most marks (50%) and §18.1 Machine learning the least (50%).
  • It is examined mostly as AO1 (Knowledge & understanding, 66%), the rest AO2 (34%) — definitions and descriptions in syllabus words score.
  • The examiner has commented on 18 of its questions — read “What the examiner said” before you practise.

Command words

Share of questions using the word (a question can use several). What each wants →

Question shapes

  • ≤ 6 mk20
  • 7–9 mk6
  • 10–12 mk3
  • 13–15 mk0
  • 16+ mk0

Average 6.2 marks a question · 41% with a figure or table · 41% with code · biggest 10 marks

Assessment objectives — how it is examined

Every part of every current-syllabus question filed under Cambridge's AO1 / AO2 / AO3 (from its command word and what it asks you to do), so you know whether this topic pays for definitions, for applying, or for judging and building.

  • AO1 Knowledge & understanding
  • AO2 Apply & analyse
  • AO3 Design, program & evaluate

Paper 3 as a whole

Paper 3SyllabusMeasured
AO1 Knowledge & understanding60%54%
AO2 Apply & analyse40%46%
AO3 Design, program & evaluate0%0%

Syllabus = Cambridge's grid; measured = the bank's current-syllabus papers.

Inside the topic — every syllabus bullet, measured

Each part of each question is filed under the bullet it examines; the numbers are per Paper 3 sitting, exactly like the topic's. Open a bullet for its own Statometer.

  • 18.1Graphs for AI#9 of 14 on Paper 3Regular · 4157% next paper2.5 marks17/32 sittings May/Jun 2026

    Set most sessions for a few marks; know the definition and one example.Syllabus: purpose and structure of a graph; searching a graph with Dijkstra's and A* algorithms

    Next Paper 3
    57%
    6 in 10
    Marks a paper
    2.5
    3% of the paper · 50% of the topic
    Asked in
    17 / 32
    Paper 3 sittings · 17 questions
    Last asked
    May/Jun 2026
    9618/33 · Q9 · 5 marks · 3-series streak
    • Asked in 17 of 32 Paper 3 sittings — about 5 papers in 10.
    • About 2.5 marks a paper (3% of Paper 3; 50% of the topic's marks across its 2 bullets).
    • Last asked May/Jun 2026 · 9618/33 · Q9 (5 marks) — in the most recent series.
    • Usually “Show” or “Explain”.
    • Biggest chunk of marks so far: 8 in May/Jun 2026 · 9618/32 · Q9.
    • It is examined mostly as AO2 (Apply & analyse, 70%), the rest AO1 (30%) — you must apply it to the given data or scenario — work it out, trace it, explain it in context.
    Last 12 sittings Steady
    Oct/Nov 24 · 31: 0 marksOct/Nov 24 · 32: 0 marksOct/Nov 24 · 33: 0 marksMay/Jun 25 · 31: 2 marksMay/Jun 25 · 32: 2 marksMay/Jun 25 · 33: 3 marksOct/Nov 25 · 31: 1 marksOct/Nov 25 · 32: 0 marksOct/Nov 25 · 33: 5 marksMay/Jun 26 · 31: 5 marksMay/Jun 26 · 32: 8 marksMay/Jun 26 · 33: 5 marks

    Assessment objectives

    • AO1 Knowledge & understanding
    • AO2 Apply & analyse
    • AO3 Design, program & evaluate
    • Show65%
    • Explain59%
    • Calculate53%
    • State29%
  • 18.1Machine learning#7 of 14 on Paper 3Regular · 4870% next paper2.6 marks21/32 sittings May/Jun 2026

    Set most sessions for a few marks; know the definition and one example.Syllabus: artificial neural networks; deep learning, machine learning and reinforcement learning (supervised and unsupervised categories); back-propagation of errors and regression

    Next Paper 3
    70%
    7 in 10
    Marks a paper
    2.6
    4% of the paper · 50% of the topic
    Asked in
    21 / 32
    Paper 3 sittings · 21 questions
    Last asked
    May/Jun 2026
    9618/33 · Q9 · 3 marks · 3-series streak
    • Asked in 21 of 32 Paper 3 sittings — about 7 papers in 10.
    • About 2.6 marks a paper (4% of Paper 3; 50% of the topic's marks across its 2 bullets).
    • Last asked May/Jun 2026 · 9618/33 · Q9 (3 marks) — in the most recent series.
    • Usually “Explain” or “Describe”: full sentences with a reason, not one-word answers.
    • Biggest chunk of marks so far: 6 in Oct/Nov 2025 · 9618/32 · Q9.
    • It is examined almost entirely as AO1 (Knowledge & understanding, 97%) — definitions and descriptions in syllabus words score.
    Last 12 sittings Steady
    Oct/Nov 24 · 31: 0 marksOct/Nov 24 · 32: 0 marksOct/Nov 24 · 33: 0 marksMay/Jun 25 · 31: 4 marksMay/Jun 25 · 32: 4 marksMay/Jun 25 · 33: 3 marksOct/Nov 25 · 31: 5 marksOct/Nov 25 · 32: 6 marksOct/Nov 25 · 33: 0 marksMay/Jun 26 · 31: 4 marksMay/Jun 26 · 32: 1 marksMay/Jun 26 · 33: 3 marks

    Assessment objectives

    • AO1 Knowledge & understanding
    • AO2 Apply & analyse
    • AO3 Design, program & evaluate
    • Explain57%
    • Describe43%
    • State29%
    • Show29%

12% of the topic's marks sit in question parts that belong to another topic (scenario questions cross sections) or that no bullet claims; they count for the topic, not for a bullet.

Marks a paper, year by year

212223242526

By exam series

  • May/Jun15/18 · 6.5 mk
  • Oct/Nov11/14 · 4.4 mk

What you need to know2syllabus §18.1

  1. 18.1Graphs for AIpurpose and structure of a graph; searching a graph with Dijkstra's and A* algorithms
  2. 18.1Machine learningartificial neural networks; deep learning, machine learning and reinforcement learning (supervised and unsupervised categories); back-propagation of errors and regression

Video lectures3ZAK's YouTube channel · play here

  • A22023374 views

  • O LevelA2Podcast2025537 views

  • O LevelA2Podcast2025226 views

Infographics2draw these the way the examiner expects · download as PNG

Graphs & Dijkstra's shortest pathA graph = nodes joined by edges, optionally weighted and directed. Dijkstra finds the cheapest path fromone node to all others.412583ABCDEShortest A → E = A-C-B-D-E, cost 1 + 2 + 5 + 3 = 11 (not the “obvious” A-B-D-E = 12).visitABCDEstart0A041C0319B0318D031811E031811Each step: pick the unvisited node with thesmallest distance, mark it visited, relaxits neighbours: if dist[u] + w < dist[v] thenupdate dist[v] and record u as previous[v].Walk previous[] back from E for the path.Representing a graphABCDEA04100B00050C02080D00003E00000matrix (left): fast lookup, O(n²) spacelist: A → [(B,4),(C,1)] — compact if sparsedirected = arrows; undirected = symmetricA* searchDijkstra + a heuristic h(n) (e.g. straight-linedistance to the goal). Expands the node with thelowest f = g (cost so far) + h. Same answer whenh never over-estimates, but far fewer nodes visited.Uses: sat-nav, packet routing, game path-finding, social graphs. Weighted graphs need Dijkstra; unweighted can use BFS.cswithzak.com

Graphs & Dijkstra's algorithm

A2
Machine learning & neural networksThree ways a program can learn from data instead of fixed rules — plus the network structure and howback-propagation trains it.Supervisedlearns from LABELLED examplesinput → known output;classification (spam?) and regression (price)Unsupervisedfinds structure in UNLABELLED dataclustering customers, anomalydetection, recommendationReinforcementlearns by trial, reward and penaltyagent acts in an environment; gameplaying, robot controlArtificial neural networkinputhiddenhiddenoutputEach node sums weighted inputs, applies an activation function,passes the result on. Weights are the “knowledge”.Deep learning = many hidden layers; needs big data + GPUs.Back-propagation: compare output with the correct answer,compute the error, push it backwards through the layers andnudge every weight to reduce it. Repeat over thousands ofexamples (epochs) until the error is small enough.Regressionfits a line/curve to data so anumeric value can be predicted(house size → price)From a scenario: labelled examples = supervised; grouping unknown data = unsupervised; reward signal = reinforcement.cswithzak.com

Machine learning & neural networks

A2

Browse all infographics →

Key terms12use these exact words in the exam

graphnodeedgeDijkstraA*heuristicsupervised learningunsupervised learningreinforcement learningneural networkback-propagationregression

Dotted terms are defined in the glossary.

Code help1referenced to the Cambridge pseudocode guide

Dijkstra on a tiny graph (adjacency matrix)

pseudocode §3 2D arrays Run in Playground
CONSTANT N = 4
CONSTANT INF = 9999
DECLARE W : ARRAY[1:4, 1:4] OF INTEGER
DECLARE Dist : ARRAY[1:4] OF INTEGER
DECLARE Visited : ARRAY[1:4] OF BOOLEAN
DECLARE i, j, U, Best : INTEGER
FOR i 1 TO N
FOR j 1 TO N
W[i, j] INF
NEXT j
Dist[i] INF
Visited[i] FALSE
NEXT i
W[1,2] 4
W[1,3] 1
W[3,2] 2
W[2,4] 5
W[3,4] 8
Dist[1] 0
FOR i 1 TO N
Best INF
FOR j 1 TO N
IF NOT Visited[j] AND Dist[j] < Best THEN
Best Dist[j]
U j
ENDIF
NEXT j
Visited[U] TRUE
FOR j 1 TO N
IF W[U, j] < INF AND Dist[U] + W[U, j] < Dist[j] THEN
Dist[j] Dist[U] + W[U, j]
ENDIF
NEXT j
NEXT i
FOR i 1 TO N
OUTPUT "Node ", i, " distance ", Dist[i]
NEXT i

Playground examples1runnable program for this topic

  • Graph as an adjacency matrix — breadth-first search

    A 2D BOOLEAN array holds the edges; BFS uses a queue to visit every node reachable from node 1.

    A2PseudocodeTheory in code 9618 §18.1
    Run

Declarative Lab2Prolog knowledge bases with sample queries

  • Routes through a network

    Is there a path from one station to another? Directed edges plus a recursive route/2. Try a start with no outgoing edge.

    A2Recursive rules 9618 §20.1
    Run
  • Exam: classifying animals

    Rules chained on rules — the expert-system style question. Ask what a given animal is and which animals are mammals.

    A2Exam-style scenarios 9618 §20.1, §18 AI
    Run

Test yourself

Ready to check you know it?

Every round is a fresh random draw, weak cards come back until you get them right, and past-paper questions come with their mark schemes. Marks earn XP on your dashboard.

Enroll nowOnline classes