Learn
6.3.3 | OPERATIONS AND CONTROLS OF AI
01 | THE TWO SYSTEMS IN THIS SYLLABUS
The specification limits this section’s basic AI operation to expert systems and machine learning.
An expert system applies stored knowledge and rules to infer a conclusion. Machine learning allows a program to automatically adapt its processes or data using training information.
The course title includes “controls”. Here, the main assessed content is how these systems operate and which components they use. Checking inputs, testing and human oversight provide useful context for controlling their use.
02 | THE FOUR EXPERT-SYSTEM COMPONENTS
| Component | Main role |
|---|---|
| Knowledge base | Store relevant facts and expert knowledge about the domain |
| Rule base | Store rules linking conditions to conclusions or actions |
| Inference engine | Apply appropriate rules to the available information to infer results |
| Interface | Let the user supply information and receive questions or results |
Implementations can organise their information differently, but these are the four distinct components named in the syllabus.
The inference engine uses the knowledge and rules; it is not simply another name for the knowledge base.
03 | SEE HOW THE COMPONENTS CONNECT
USER INTERFACE
Questions, user answers and displayed recommendations
↕ Supply information / return results
INFERENCE ENGINE
Check information and apply relevant rules
↕ Consult knowledge and rules
KNOWLEDGE BASE
Facts and domain knowledge
RULE BASE
Conditions and conclusions
The interface does not decide the answer merely because it displays it. The inference engine performs the reasoning using the available knowledge and rules.
04 | WORKED EXPERT SYSTEM: PRINTER TROUBLESHOOTING
Use an invented expert system that recommends a first check, rather than claiming to identify every fault.
| Component | Teaching example |
|---|---|
| Knowledge base | Knowledge about printer power, paper and indicators |
| Rule base | R1: IF power absent THEN recommend checking power<br>R2: IF power present AND paper absent THEN recommend checking paper tray |
| Interface | Ask whether power is present and paper is loaded |
| Inference engine | Check the answers against the rule conditions |
The user answers: power is present; paper is absent. R1 does not match. Both conditions in R2 match, so the engine infers the recommendation to check the paper tray.
The interface displays the recommendation. It does not prove the tray is the only possible fault.
05 | TRACE THE REASONING
- Obtain relevant information.
The interface gathers answers about the current problem.
- Make the information available for reasoning.
The system combines current facts with its stored domain knowledge.
- Apply relevant rules.
The inference engine checks which conditions match.
- Infer a conclusion or ask for more information.
The available facts may support a recommendation or require another question.
- Present the result.
The user receives it through the interface.
Different systems can reason in different orders. The key syllabus idea is that an inference engine applies knowledge and rules, rather than selecting an unsupported answer at random.
06 | EXPERT-SYSTEM LIMITS
Wrong input: an incorrect user answer can cause the wrong rule to be applied.
Missing knowledge: an unusual fault may not be covered.
Outdated rules: a changed product or process may require the knowledge and rules to be updated.
A fixed expert system does not necessarily learn automatically when a user disagrees. People can review its behaviour and amend its information.
07 | MACHINE LEARNING: ADAPT FROM DATA
The specification describes machine learning as a program having the ability to automatically adapt its own processes and/or data.
In a common example, training data is used to adjust a model that will later classify an input or predict an outcome.
People still choose the task, data, model and training process. Automatic adaptation does not mean the system creates its own purpose or understands the world like a person.
Unlike a simple rule-based example, the programmer need not explicitly write a separate rule for every input pattern.
08 | WORKED MACHINE LEARNING: MATERIAL IMAGES
A school project aims to classify images into paper, metal and plastic categories. This is a teaching scenario, not a claim that every item can be identified from an image.
| Stage | What happens |
|---|---|
| Collect suitable examples | Obtain representative images with appropriate category labels |
| Prepare data | Check labels and use a consistent suitable representation |
| Train | The training process adjusts the model using the examples |
| Evaluate | Check performance on suitable examples not used to fit the model |
| Use / inference | Provide a new image and obtain a predicted category |
| Review | Investigate errors and decide whether further training or design changes are needed |
The predicted category can be wrong, especially for unclear, unusual or poorly represented examples.
09 | TRAINING IS NOT THE SAME AS USE
TRAINING
Examples → adaptation process → adjusted model
INFERENCE
New input → trained model → prediction
Training changes the model. Inference uses it. A deployed model can remain fixed while producing different outputs for different inputs.
Some systems support ongoing learning, but not every model changes on every request. A new answer alone is not evidence that learning occurred.
10 | WHY TEST WITH DIFFERENT EXAMPLES?
A model may fit familiar examples well but perform poorly on new ones. Generalisation concerns how the learned relationships work beyond the fitting examples.
Using suitable evaluation data not used to fit the model gives a better check than only reusing training examples.
Example: a classifier trained mainly on clean objects against plain backgrounds might struggle with crushed objects or cluttered scenes.
More examples are useful only when their quality and relevance support the task. Incorrect labels or unrepresentative data can undermine performance.
11 | COMPARE OPERATION
| Point | Expert-system example | Machine-learning example |
|---|---|---|
| Main basis | Explicit domain knowledge and rules | Model adapted from training data |
| Result production | Inference engine applies rules to facts | Trained model maps an input to a predicted result |
| Updating | People may change knowledge and rules | Training automatically adapts the model through a chosen process |
| Possible fault | Wrong facts or missing/outdated rules | Poor representation, labels or generalisation |
| Continuous learning | Not required | Not required during every use |
Real systems may combine both approaches. For the course, explain each basic method separately before comparing them.
12 | CHECKING AND HUMAN OVERSIGHT
Useful checks include validating inputs, testing representative cases, reviewing errors and limiting use to the system’s demonstrated capabilities.
Human oversight matters when a result has significant consequences. A recommendation or prediction should not be accepted simply because it is produced by AI.
When the system is uncertain or the case falls outside its intended scope, a suitable design can request more information or refer the decision to a person.
These controls support appropriate use; they do not guarantee every result is correct or unbiased.
13 | COMMON MISTAKES
- The knowledge base applies the rules: that is the inference engine’s role.
- The interface is the reasoning engine: it handles interaction.
- Machine learning means writing every pattern rule manually: training adapts a model through data.
- Every expert system learns automatically: fixed rules can stay unchanged.
- Inference always retrains the model: using and adapting it are different stages.
- Training success proves new-case accuracy: generalisation needs evaluation.
- AI output is a guaranteed fact: it depends on information, method and context.