Loading…
Free A-Level Computer Science guide
Cambridge 9618 gives AI both an AS applications/impact context and an A Level technical treatment. Explain what the system learns, what evidence trains it and how its output is evaluated.
Content owner: Michael Print · Written for A-Level learners · Checked against official specifications
Supervised learning uses labelled examples to learn a mapping; unsupervised learning seeks patterns without supplied target labels; reinforcement learning learns action choices using rewards from interaction. Deep learning uses neural networks with multiple learned layers and can support different learning setups.
Training changes model parameters to improve a defined objective. Performance on training data alone is insufficient: unseen evaluation data, bias, costs and human consequences matter. An AI output is not automatically accurate because it was computed.
Cambridge International 9618 · 9618 (2027–2029): 7.1 AI applications/impact; 18.1 graphs, neural networks, learning categories, backpropagation and regression
Useful foundations
By the end, you should be able to
In 2027–2029 version 2, AS section 7.1 covers AI applications and social, economic and environmental impacts. The checked 2026 version 2 retains these requirements.
A Level section 18.1 adds A*/Dijkstra graph search, neural networks, machine/deep/reinforcement learning, supervised/unsupervised categories, backpropagation and regression. This technical scope extends beyond OCR’s ethical discussion of AI.
Cambridge explicitly exempts writing graph setup/access/search algorithms for 18.1. Understand graphs and perform the named searches. Shared pathfinding reasoning is useful; another board’s graph-code assessment requirements must not be imported.
Learning methods use different training evidence. Match that evidence to the task before selecting a model, and decide how to measure success on unseen examples.
Supervised training reduces differences between predictions and supplied targets. Labels can be expensive or biased, and a training sample can omit rare cases. An unsupervised cluster indicates similarity, rather than a verified fault diagnosis.
A reinforcement agent seeks a policy improving expected reward. Reward design, exploration and costly or unsafe trials limit practical use; a useful reward must reflect the behaviour actually wanted.
Deep learning uses multiple learned layers that can build useful representations, such as image features. It does not exclude supervised learning: a deep network can learn from labelled images.
Larger networks can require substantial data and computation. A smaller, interpretable model may suit a simple numerical relationship better. A named AI application does not establish that AI is the best approach.
A neuron weights input values, adds a bias and applies an activation function. Connected layers pass outputs forward; the learned parameters determine the network’s mapping.
With inputs 2 and 1, weights 0.4 and −0.2, and bias 0.1, the weighted sum is 0.4×2−0.2×1+0.1 = 0.7. ReLU is defined here as max(0,z), giving output 0.7.
Backpropagation uses the chain rule to calculate each parameter’s contribution to loss through the layers. An optimisation method then updates parameters; backpropagation itself does not copy the desired answer into every weight.
The learning rate controls update size. Too large may destabilise learning; too small may slow progress. One update does not guarantee a perfect prediction.
Predict
Pass training inputs forward through the network using the current weights and biases.
Measure loss
Compare the prediction with its supplied target using the chosen loss function.
Backpropagate
Calculate how changes to each parameter would affect the loss.
Update
Use an optimisation method and learning rate to change the parameters, then repeat with training examples.
Improvement on training examples still needs independent evaluation on unseen data.
Linear regression predicts y using y-hat = w×x+b. Fitting chooses w and b to reduce training loss, rather than treating the first guessed line as truth.
Mean squared error averages squared prediction errors, strongly penalising large errors. Regression is a prediction method; a fitted relationship is not evidence that one measured variable causes the other.
Use separate held-out validation/test data. Repeatedly choosing model changes from test results leaks information into development; a final untouched test provides stronger evidence.
Overfitting captures training-specific patterns but performs poorly on new inputs. Check representative sampling, class imbalance and false-positive/false-negative consequences. High overall accuracy can hide poor performance for a small important group.
An automated fault detector can reduce manual work and catch faults early; false alarms interrupt production and missed faults create risk. Workers may gain tasks or face displacement; data collection can expose personal information.
Training and deployment consume energy. Consider benefits, error costs and the roles of oversight, appeal, staff training and proportionate data use. A reliable fixed threshold may meet a simple task with lower complexity.
Graphs represent states or locations as vertices, possible moves as edges and costs as weights. Dijkstra expands minimum cost-so-far choices for non-negative weights.
A* prioritises estimated total f=g+h: travelled cost plus a heuristic. A suitable non-overestimating heuristic supports shortest-path guarantees under the algorithm’s stated rules. The worked trace below shows repeated cost updates without graph implementation code.
Worked example
| x | Target y | Old prediction | Old error | New prediction | New error |
|---|---|---|---|---|---|
| 1 | 4 | 3 | -1 | 3.8 | -0.2 |
| 2 | 7 | 5 | -2 | 6.3 | -0.7 |
points = [(1.0, 4.0), (2.0, 7.0)]
w, b = 2.0, 1.0
errors = [w * x + b - y for x, y in points]
loss = sum(error ** 2 for error in errors) / len(points)
grad_w = 2 * sum(error * x for error, (x, y) in zip(errors, points)) / len(points)
grad_b = 2 * sum(errors) / len(points)
w -= 0.1 * grad_w
b -= 0.1 * grad_b
new_loss = sum((w * x + b - y) ** 2 for x, y in points) / len(points)
print(f"before: {loss:.3f}")
print(f"w={w:.1f}, b={b:.1f}")
print(f"after: {new_loss:.3f}")Worked example
| Expanded | A | B | G |
|---|---|---|---|
| S | 2 | 5 | ∞ |
| A | 2 | 3 | 8 |
| B | 2 | 3 | 5 |
| G | 2 | 3 | 5 |
Worked example
7 original questions total 24 marks. Attempt each before opening the independently written indicative marking guidance.
Match three tasks to supervised, unsupervised or reinforcement learning: labelled disease images, grouping unlabelled activity logs, and an agent choosing actions from rewards.
Indicative answer
Explain a neuron's weighted input, bias and activation, then calculate the output of this neuron.
Bias is 0.1. Add the bias to the sum of the weighted inputs to obtain z, then use ReLU activation, defined as max(0,z).
| Input | Value | Weight |
|---|---|---|
| First | 1 | 0.4 |
| Second | 3 | -0.2 |
Indicative answer
Explain the roles of loss, backpropagation and learning rate during training.
Indicative answer
A regression model predicts y=3x+1. For points (1,5) and (2,6), calculate predictions and mean squared error.
Indicative answer
A model achieves perfect accuracy on its training data. Give two reasons this is insufficient and one improvement to evaluation.
Indicative answer
Evaluate an AI recruitment filter with one benefit, two risks and one practical safeguard.
Indicative answer
The graph below uses a changed S→B cost of 1. Give the shortest route/cost, calculate B's initial A* priority using h(B)=2, and assess the old h(S)=5.
All edges are directed as listed. Start at S; the goal is G. The old heuristics are h(S)=5, h(A)=3, h(B)=2 and h(G)=0. For A*, use f = g + h, where g is the cost travelled from S.
| From | To | Cost |
|---|---|---|
| S | A | 2 |
| S | B | 1 |
| A | B | 1 |
| A | G | 6 |
| B | G | 2 |
Indicative answer
This guide addresses Cambridge International 9618 9618 (2027–2029): 7.1 AI applications/impact; 18.1 graphs, neural networks, learning categories, backpropagation and regression. Check your examination year and the complete specification for the assessment scope.
These are independently written explanations and practice questions. CompSciTutoring.co.uk is not affiliated with or endorsed by an examination board. The marking guidance is indicative; always check the syllabus for your examination year.
Found an error? Email a correction.
Use the free examples first, then explore one-to-one tuition if guided practice would help.