Intelligent Systems

Decision Support Systems

🎯 (EC-12-01, EC-12-04)

📌 Investigate applications of decision support systems

A decision support system (DSS) is an information system that helps people make better decisions by combining data, a model of the problem and an interface the decision-maker can question. The key word is support: a DSS presents options, forecasts and consequences, and a person still owns the decision. NESA groups it with the intelligent systems that gather and analyse data and produce targeted output.

What a DSS is made of

• Data: internal records (sales, sensor readings, enrolments) and external sources (weather, prices, census data), which must be accurate and current.
• Model: the rules, formulas or statistical methods that turn data into a recommendation, such as a forecast, a score or a ranking.
• Interface: dashboards, reports and "what if" controls that let the user change an assumption and see the result straight away.

Applications across enterprises

Investigating applications means asking what decision is being supported, which data feeds it and who is affected. The table compares six realistic uses.

Benefits and limits

A DSS can handle more data than a person can, apply the same rules every time, and show the effect of a change before it is made. Its limits are just as important. A recommendation is only as good as its data and model, so out-of-date or biased data produces confident but wrong advice. Users may also trust the screen too much (automation bias) and stop questioning it. The legal and ethical duties stay with the enterprise: a bank using a scoring model still has to be able to explain a refusal, and the Privacy Act 1988 applies to the personal information the model uses.

Categories of Decision-Making

🎯 (EC-12-01, EC-12-02)

📌 Describe categories of decision-making within decision support systems

• an infinite set of instructions where judgement may be required (unstructured)
• use of a specified set of finite instructions (semi-structured)
• automated decisions (structured)

NESA describes decision-making in a DSS by how much of the decision can be written down as instructions. At one end a person must use judgement because no complete set of instructions exists; at the other end the instructions are so complete that the system can decide alone. Between them is the large middle ground where the system follows defined steps but a person confirms the outcome.

An infinite set of instructions where judgement may be required (unstructured)

An unstructured decision has no complete procedure. The situation is new, the factors are many and unclear, and different sensible people could reach different answers. The DSS cannot decide, but it can gather data, chart trends and run "what if" models so the decision-maker is better informed. Examples: whether a coffee roaster should open a second shop in another town; how a school should respond to a sudden drop in Year 9 attendance; choosing which new product line a start-up should build.

Use of a specified set of finite instructions (semi-structured)

A semi-structured decision has a defined procedure for most of the work, but some part still needs judgement or approval. The DSS carries out the finite steps (calculate a score, apply a policy, compare options) and hands a recommendation to a person. Examples: a loan officer receives a risk score built from set rules but may approve an exception for a long-standing customer; a supermarket manager reviews a suggested weekly order and lifts it before a local festival; a hospital coordinator accepts or changes a proposed roster.

Automated decisions (structured)

A structured decision is fully specified, so the system makes it automatically and no one is needed at the moment of decision. Examples: a warehouse system that reorders an item when stock falls below a set level; a thermostat that switches a heater on below 18 degrees Celsius; traffic signals that follow a timing plan; an online form that rejects a card number that fails a check-digit test. A person designed and tests the rules beforehand, and reviews the results afterwards.

Applications of Expert Systems

🎯 (EC-12-01)

📌 Investigate common applications of expert systems

• diagnosis
• monitoring
• process control
• scheduling and planning

An expert system is a program that captures the knowledge of human experts as facts and rules, and uses them to reach conclusions or advice about a narrow topic. It suits problems where expertise is scarce, decisions must be made quickly or consistently, and the knowledge can be written as rules. The four applications below cover most of what the syllabus expects; investigate each by asking what goes in, what comes out and where a person is still needed.

Diagnosis

A diagnostic expert system takes symptoms and observations and works out the most likely cause. A vehicle workshop tool reads fault codes and sensor readings and suggests which part to test first; an IT help desk system asks about error messages and lights until it finds the fault; clinical decision support tools suggest possible conditions to a doctor, who confirms the diagnosis. Rules such as IF the power light is off THEN check the cable are cheap to write and easy to explain, but the system only knows the faults its rules cover.

Monitoring

Monitoring systems compare a stream of readings against what is expected and raise an alert when something is wrong. Examples include an intensive care monitor that warns when heart rate and oxygen readings together suggest a patient is worsening, a network security tool that flags unusual login patterns, and a cold-chain system that notices a refrigerated truck warming up. The value is constant attention: the system does not tire. The risk is alert fatigue if too many false alarms teach staff to ignore it.

Process control

Process control expert systems adjust equipment automatically to keep a process on target. A dairy factory might change pasteurising temperature and flow rate as milk conditions change, and a smart greenhouse opens vents and starts fans when temperature and humidity move outside set ranges. The system reads sensors, applies rules and drives actuators without waiting for a person, so testing and safe limits are essential. It is also a core idea of Industry 4.0 manufacturing.

Scheduling and planning

Scheduling systems fit tasks, people and resources together under rules and constraints. Examples: a school timetable that avoids clashes for teachers and rooms; a hospital roster that respects rest periods; a freight planner that sequences deliveries within driver-hour limits; an airline crew planner. Rules express the constraints (IF a driver has worked 5 hours THEN a break is needed), and the system searches for a plan that satisfies as many as possible. Planners still review the result, because real life brings sick leave and road closures.

Key Features of an Expert System

🎯 (EC-12-02, EC-12-08)

📌 Describe key features of an expert system

• knowledge base
• inference engine, including forward chaining and backward chaining
• user interface (UI)

Every expert system, from a small help-desk advisor to a large industrial controller, is built from the same three parts: a knowledge base that stores what is known, an inference engine that reasons with it, and a user interface that talks to the person using it. Keeping the knowledge separate from the reasoning is what makes the design powerful: an expert can update a rule without anyone rewriting the program.

Knowledge base

The knowledge base holds facts and rules. A fact is a statement taken as true, such as soil_wet or "the pot has no drainage holes". A rule is an IF–THEN statement joining facts to a conclusion: IF leaves_yellow AND soil_wet THEN overwatered. The IF part is the condition (or antecedent) and the THEN part is the conclusion (or consequent). Conditions can be joined with AND (all must be true) and OR (either may be true). A conclusion can itself become a fact that triggers further rules, which is how the system chains several small steps into an answer.

The knowledge comes from human experts through a knowledge engineer, who interviews them, reads manuals and turns what they know into rules. That is called knowledge acquisition, and it is usually the slowest and most expensive stage. NESA's course specifications expect expert system software to let you develop a set of IF–THEN rules and add, remove and edit them, which is what the demo below lets you try.

Inference engine, including forward chaining and backward chaining

The inference engine is the reasoning program. It matches the facts in working memory against the rules, decides which rule to apply, and adds new conclusions. Two reasoning strategies are examined in this course. In NESA's glossary, forward chaining evaluates existing data to develop new data to make a decision, and backward chaining uses the goal or output to track back and establish the steps taken.

Both are worked below on one small rule base for a plant-care advisor.

Case: the user reports three facts: leaves_yellow, soil_wet and no_drainage.

Forward chaining: from the facts to a conclusion

Forward chaining is data-driven. The engine looks for a rule whose conditions are all in working memory, fires it, adds the conclusion, and repeats until no new rule can fire. When more than one rule could fire, a conflict-resolution rule chooses; here the lowest-numbered rule goes first.

The advice is to reduce watering and repot into a pot with drainage. Forward chaining discovered everything that follows from the facts, including a conclusion nobody asked for. That suits monitoring and process control, where data arrives and the system must react.

Backward chaining: from a goal back to the facts

Backward chaining is goal-driven. The engine starts with a hypothesis, finds the rules that could prove it, and treats each condition as a new sub-goal, until it reaches facts it can check or ask the user about. It suits diagnosis, where you want to confirm one possibility. The goal here is repot_with_drainage.

If the goal were move_to_shade, the engine would try R8, then R3, and ask whether brown_edges is true. The answer is no, so R3 fails, R8 fails and the goal cannot be proved. Backward chaining asks only the questions that matter to the goal, which keeps the dialogue short. The chain of rules that succeeded is also the explanation of how the answer was reached.

User interface (UI)

The user interface is how people supply facts and receive advice. A good one asks plain-language questions in a sensible order (backward chaining helps here), accepts answers quickly by buttons or drop-downs, and shows the conclusion with its recommended action. Most expert systems add an explanation facility that answers "why are you asking me that?" and "how did you reach that conclusion?" by showing the rules used. NESA's specification says expert system software should display the rules used to reach a conclusion in text and as a decision tree, as the demo's explanation tree does. Explanations build trust and let an expert spot a wrong rule. A poorly designed interface wastes a good knowledge base, so the design principles from Year 11 apply: consistency, clear labels, and helpful feedback.

From Rules to Probability

🎯 (EC-12-06)

📌 Research factors that have allowed expert systems to advance from rule-based systems to integrate probability

• processing power
• data availability and access
• hardware costs
• neural networks
• webometrics

Early expert systems in the 1970s and 1980s were rule-based: every conclusion followed from rules written by hand, and each rule was either true or false. That worked for narrow problems but had three weaknesses. Writing rules was slow because experts find it hard to put their judgement into words. The systems were brittle, giving no answer or a wrong one when a case fell outside the rules. And they handled uncertainty poorly, because real evidence is rarely a clean yes or no. Modern intelligent systems add probability: they say how likely a conclusion is, and they learn patterns from data instead of relying only on hand-written rules. Researching why this became possible points to five linked factors.

Processing power

Probabilistic methods and neural networks need huge numbers of calculations. Faster processors, and especially graphics processing units (GPUs), which do many simple calculations in parallel, made it practical to train models that once took months. Cloud computing lets an enterprise rent that power by the hour instead of owning it. More power also lets an inference engine weigh thousands of uncertain pieces of evidence in the time a person waits for an answer.

Data availability and access

A system can only learn patterns from data, so the growth of digital data changed what was possible. Websites, smartphones, cameras, IoT sensors and digital records now produce large datasets (big data) as a side effect of daily life, and open data portals and cloud storage make them easier to reach. More data reveals patterns that a handful of experts would never state as rules. The limits are quality and fairness: data must be accurate, lawfully collected and representative, or the model inherits its gaps and biases.

Hardware costs

Storage, memory, processors and sensors have become far cheaper over time, so it is now affordable to collect data continuously and keep it. A small business can add temperature sensors that cost little, send data to cheap cloud storage and run analysis it could not have paid for a generation ago. Lower costs also spread intelligent systems to phones, cars and household devices, which produces still more data.

Neural networks

An artificial neural network is a set of connected layers of simple units. Each unit combines its inputs using weights and passes a value on; training adjusts the weights until the outputs match known examples. Instead of an expert writing a rule such as "IF the link is untrusted THEN spam", the network learns from thousands of labelled emails which patterns matter. Its output is typically a probability, for example a 0.93 likelihood that a message is spam. Neural networks became practical once processing power, data and lower hardware costs came together. Their weakness is that the learned weights are hard to explain, unlike an IF–THEN rule.

Webometrics

Webometrics is the quantitative study of the web: measuring and analysing the structure, content and use of web resources, such as the hyperlinks between pages, the number and type of pages on a site, and how people search and click. It draws on the methods used to measure other bodies of information. Search engines apply the same idea. Google's original PageRank method treated a link to a page as a kind of vote and estimated the probability that someone following links at random would reach it. Webometrics is a source of data (what people link to, search for and read) and a way of estimating how likely a page is to be relevant, which moves web ranking from fixed rules to probability.

Changing Needs and Expert Systems

🎯 (EC-12-01, EC-12-06)

📌 Describe how people’s changing needs have shaped the proliferation of expert systems

• the Internet of Things (IoT)
• Internet of Me (IoMe)
• Industry 4.0

Expert systems spread because people and enterprises wanted things they could not get from paper records and human effort alone: faster answers, advice tailored to one person, less waste and safer workplaces. Three ideas in the syllabus describe how those needs changed. Each created demand for systems that sense, decide and act with little human involvement.

The Internet of Things (IoT)

NESA's glossary describes the Internet of Things as smart devices and objects that send and receive information using the internet and communication infrastructure. As sensors and network connections became cheap, homes, farms and factories wanted to monitor and control equipment from anywhere. A connected soil probe or a cold-room sensor is only useful if something interprets its readings. An expert system is a natural fit: it applies rules to a constant stream of readings and acts, so IoT created demand for systems that can decide without a person watching a screen.

Internet of Me (IoMe)

NESA describes the Internet of Me as technology that connects our minds and bodies with the online world, turning biological and cognitive life into streams of data that can be monitored, shared and shaped. Fitness bands, smartwatches, sleep trackers and glucose monitors are examples. People want personal advice such as "you have not moved much today" or "your heart rate is unusually high", and health providers want to spot problems earlier. Intelligent systems supply that personalisation, but the data is highly sensitive: it is covered by the Privacy Act 1988 as health information where it identifies a person, and it raises questions about who owns and profits from it.

Industry 4.0

Industry 4.0 describes a fourth wave of industrial change in which factories and supply chains become connected, data-driven and increasingly automated, combining IoT sensors, cloud computing, data analytics and AI. A smart factory can monitor every machine, predict a failure before it happens, adjust a process automatically and reorder parts. Manufacturers wanted less downtime, more flexibility, lower waste and products made to order. Expert systems and machine learning are the decision-making layer that turns sensor data into action.

Intelligent Agents and Search Engines

🎯 (EC-12-04, EC-12-06)

📌 Describe the operation of simplistic and complex intelligent agents used by search engines, including predictive search strings and the use of voice assistants to initiate interaction

An intelligent agent is a program that perceives its environment through inputs, decides what to do, and acts to reach a goal. A search engine is a good case study because it uses agents of very different sophistication, from simple ones that follow fixed rules to complex ones that learn.

How a search engine works

A search engine has three stages. Crawling: software agents called crawlers follow links from page to page and collect content. Indexing: the content is analysed and stored so it can be found quickly. Ranking: when someone searches, the engine matches the query to the index and orders the results by relevance. Agents work at each stage.

Simplistic and complex agents

Predictive search strings

When you type "how to cook" and the box offers "how to cook rice" before you finish, an agent is predicting the rest of your search string. It estimates the most likely completions from how often people search for them, then adjusts using signals such as your location, language, recent searches and trending topics. Each suggestion is a probability estimate, not a fixed answer. This saves time and helps people with spelling, but it also has costs. Suggestions can reflect the biases of the people whose searches trained them, can steer attention towards popular topics, and depend on the collection of personal search history, which raises privacy questions. Search providers filter certain suggestions, and users can usually manage or delete their history.

Voice assistants

A voice assistant lets a person start an interaction by speaking. A small program on the device listens for a wake word, and only after hearing it does the rest of the request get processed. The steps are: speech recognition turns the audio into text, language understanding works out the intent and details, an agent acts (search, answer a question, set a timer or control a smart device), and the reply is spoken or displayed. Some processing happens on the device and some in the cloud, depending on the product. Voice suits people who have their hands full or find typing hard, and it supports accessibility. The concerns are accidental activation, the storage of recordings, and mistakes with accents or background noise.

Inference Engine Techniques

🎯 (EC-12-02, EC-12-08)

📌 Compare techniques used by different inference engines

• truth maintenance
• hypothetical reasoning
• heuristic knowledge and fuzzy logic
• ontology classification

Forward and backward chaining decide the direction of reasoning. Real problems need more: facts change, evidence is uncertain, and knowledge is organised in hierarchies. Inference engines therefore use additional techniques. Comparing them means asking what problem each solves, how it works, and what it costs.

Truth maintenance

A truth maintenance system records why each conclusion is believed, meaning which facts and rules it depended on. When a fact is later changed or withdrawn, the system automatically retracts every conclusion that relied on it, and restores alternatives. Without it, an engine would keep acting on stale conclusions. Example: in the plant advisor, overwatered was concluded from soil_wet. If the user corrects the answer to "the soil is actually dry", truth maintenance withdraws overwatered, root_rot_risk and the advice to repot.

Hypothetical reasoning

Hypothetical reasoning explores "what if" by making an assumption and following its consequences in a separate, temporary context, without changing the real facts. If the assumption leads to a contradiction it is discarded; if not it can be adopted. Example: a roster system assumes a nurse works Thursday night, checks the rest-period rules, finds a clash on Friday and abandons that option. It is also the idea behind the what-if tools in a decision support system.

Heuristic knowledge and fuzzy logic

A heuristic is a rule of thumb from experience that finds a good-enough answer quickly, without guaranteeing the best one. A mechanic's rule "check the battery first" is heuristic knowledge. It lets an engine skip unlikely paths, which matters when there are too many possibilities to check.

Fuzzy logic deals with vagueness. Ordinary logic says a temperature is hot or not hot. Fuzzy logic gives it a degree of membership from 0 to 1 in each set such as cold, warm and hot, which suits words people actually use. The worked example follows.

Worked example: a greenhouse fan. The membership functions in the chart are: Warm rises from 0 at 10 °C to 1 at 20 °C, then falls to 0 at 30 °C; Hot rises from 0 at 20 °C to 1 at 30 °C. The rules are: IF temperature is Warm THEN fan speed is Medium (50%); IF temperature is Hot THEN fan speed is High (100%). The temperature is 24 °C.

1. Fuzzify. Warm at 24 °C = (30 − 24) ÷ (30 − 20) = 0.6. Hot = (24 − 20) ÷ (30 − 20) = 0.4. Cold = 0.
2. Apply the rules. The Warm rule is true to degree 0.6 and the Hot rule to degree 0.4.
3. Combine (defuzzify). Use a weighted average of the fan speeds: (0.6 × 50 + 0.4 × 100) ÷ (0.6 + 0.4) = (30 + 40) ÷ 1 = 70%.

A crisp rule such as "IF temperature > 25 THEN fan High" would give a sudden jump between 24.9 and 25.1 degrees. The fuzzy fan speed rises smoothly as the greenhouse warms, which is gentler on equipment and closer to how an expert would describe it. (Real fuzzy controllers use several ways of combining, and this weighted average is a simplified one.)

Ontology classification

An ontology is a formal description of the concepts in a domain and the relationships between them, such as is a kind of and is part of. Classification places a new item into the hierarchy, so the engine can infer properties it inherits. Example: a product-recommendation system knows that a "trail running shoe" is a kind of "running shoe", which is a kind of "footwear". A rule written for running shoes therefore applies to trail shoes without being repeated. Ontologies are used in medical coding, product catalogues and the semantic web. Their cost is the effort to build and agree on the hierarchy.

These techniques are combined rather than chosen once. A plant-care advisor might use forward chaining with truth maintenance, a fuzzy rule for watering amounts, and an ontology so that "fern" and "orchid" inherit the rules for "indoor plants".

Hardware in Intelligent Systems

🎯 (EC-12-06, EC-12-08)

📌 Investigate the hardware used in an intelligent system

• biometrics
• haptics
• touch and gesture
• virtual and augmented reality (VR/AR)
• voice and sound
• microcontrollers
• sensors, actuators and motors

Intelligent systems need hardware to sense the world, process what they sense and act on the result. Investigating the hardware means asking what each device does for the system, where it sits between input and output, and what limits it. The first figure shows the overall pattern of input, processing and output, with feedback closing the loop.

Biometrics

Biometric hardware measures a body feature to identify or verify a person: fingerprint readers, face recognition cameras, iris scanners and voice recognition. NESA's glossary uses biometric to describe data whose source is human or biological. They suit access control and phone unlock because the feature is hard to forget. The risks are false matches and misses, bias in recognising some groups, and permanence: a stolen password can be changed, a fingerprint cannot. Biometric data is sensitive information under the Privacy Act 1988.

Haptics

Haptic technology relates to the sense of touch, most often as an output measured by feel. NESA's glossary describes haptic as output from a device that can be felt. A phone vibrating, a game controller rumbling and a surgical simulator that resists the trainee's hand are examples. Haptics gives feedback without a screen, which helps people who cannot look at a display and adds realism in training and virtual reality.

Touch and gesture

Touchscreens detect a finger's position, and multi-touch recognises pinch and swipe. Gesture input goes further: cameras or motion sensors track hand and body movements without contact, useful in a kitchen, an operating theatre or a vehicle where touching a screen is awkward. The design task is to make gestures learnable and to avoid accidental triggering. Both are natural interfaces, since people already know how to point and wave.

Virtual and augmented reality (VR/AR)

NESA describes virtual reality as a simulated environment that places the user inside an experience, and augmented reality as replicating or overlaying extra information about the real world using computer-generated data such as GPS, sound, video and images. VR uses headsets with motion tracking and often handheld controllers. AR uses a phone, tablet or glasses, as in a heads-up display in a car. Intelligent systems supply the content: they track where the user looks and what they do, and respond. Applications include surgical and pilot training and inspection of equipment overlaid with live sensor data. Limits are cost, motion sickness for some users and the need for accurate tracking.

Voice and sound

Microphones capture speech and sound, and speakers give spoken output. Beyond voice assistants, sound sensors can detect a machine running roughly, glass breaking or a baby crying. Speech input works hands-free and supports accessibility, but accuracy falls in noisy places and with unfamiliar accents, and always-on microphones raise privacy concerns.

Microcontrollers

NESA defines a microcontroller as a small computer built for specific tasks, such as managing a car engine or a microwave's display, which processes data input by users (a keypad) or by sensors (light). A single chip typically holds a processor, memory and input and output pins. Microcontrollers are cheap, use little power and run one program continuously, which makes them the brain of most IoT devices. Their small memory limits what they can do locally, so heavier analysis is often sent to a server or the cloud. Small development boards are widely used for student projects.

Sensors, actuators and motors

A sensor detects and responds to some type of input from the physical environment, such as temperature, light, motion, moisture, pressure, sound or location. An actuator is the opposite: it converts a control signal into a physical action, such as opening a valve, switching a relay or moving a vent. A motor is an actuator that produces rotation or linear movement: a pump motor, a servo that moves a robot arm, a conveyor drive. The sensors supply facts to the rules and the actuators carry out the conclusions.

Computational Thinking in Design

🎯 (EC-12-08)

📌 Explore how computational thinking can be integrated into the design and development of an intelligent system

• decomposition
• pattern recognition
• abstraction
• algorithms

NESA's glossary defines computational thinking as a process in which a problem is analysed and solved so that a human, machine or computer can implement the solution: organising data logically, breaking a problem into parts, interpreting patterns, and designing and implementing algorithms. When designing an intelligent system it is a working method, not just a list of terms. This section follows one problem, a smart greenhouse controller for a plant nursery, through the four steps.

Decomposition

Decomposition breaks a large problem into smaller, more manageable ones that can be solved separately, by different people if needed. "Control the greenhouse" is too large to design at once. Splitting it into sense, decide, act and report, then splitting those again, gives parts that can be built and tested one at a time. The same split becomes the processes in a data flow diagram and the modules in the code.

Pattern recognition

Pattern recognition means spotting similarities, trends and regularities. In the nursery the manager notices that soil dries fastest on hot afternoons and that seedlings in the north bench need water more often. Those patterns become rules ("IF temperature is above 30 AND time is after noon THEN check moisture more often"), or they are found automatically by machine learning in the sensor history. Recognising that many decisions share one structure also lets you reuse a rule: the moisture-versus-threshold logic is the same for every bench.

Abstraction

NESA describes abstraction as representing data and programs in a form similar to their meaning while hiding implementation details, temporarily ignoring detail so you can focus on a few concepts at a time. For the greenhouse, a plant is abstracted to a few values: soil moisture, temperature and growth stage. The colour of the pots and the brand of the sensor are ignored. A good abstraction includes what affects the decision and leaves out the rest, and models and diagrams are abstractions. Too little detail makes wrong decisions, and too much makes the system slow and hard to maintain.

Algorithms

An algorithm is a step-by-step procedure that solves a problem. Once the problem is decomposed, patterns found and detail abstracted, the algorithm is the exact sequence the system follows. It may be written as steps, drawn as a flowchart or coded. The watering algorithm for one bench is:

1. Read the soil moisture and the rain forecast.
2. If moisture is 40% or more, stop: no watering is needed.
3. If rain is forecast today, stop: let the rain do the work.
4. Otherwise run the pump for 10 minutes, then send a watering alert to the grower.
5. Repeat at the next reading.

The same algorithm appears as a flowchart in the next section and as a program later in this page. Sequence, selection (the two decisions) and repetition are enough for most control tasks.

Communicating Logical Processes

🎯 (EC-12-08, EC-12-11)

📌 Communicate the logical processes performed by an intelligent system by using flowcharts, data flow diagrams and infographics

Once a system's logic is designed it has to be communicated to clients, teammates and later users. Three tools suit different jobs: flowcharts show the order and decisions in a process, data flow diagrams show how data moves between people, processes and storage, and infographics present results to a general audience. The first two must use the notation in NESA's Enterprise Computing Course Specifications, which is examinable. The notation on this page is redrawn from the specifications; NESA is the source, and the official document is linked in the site footer.

Flowcharts

A flowchart represents logic and is read from top to bottom and left to right. NESA uses four symbols: an input or output parallelogram, a terminator (rounded ends, reading START and END), a process rectangle and a decision diamond. Arrows show the direction of flow, and every arrow leaving a decision is labelled Yes or No.

Data flow diagrams

A data flow diagram (DFD) models a system by the data that moves through it. NESA's notation has four symbols: a circle for a process (which uses inputs to make outputs), an open-ended rectangle for a data store (an electronic file or non-computer storage), a closed near-square rectangle for an external entity (any person, organisation or element that provides data to, or receives it from, the system) and a labelled, curved arrow for a data flow. A Level 0 diagram gives an overview of the whole system: one process, the external entities, and no data stores or internal processes. A Level 1 diagram opens that process into its main sub-processes and shows the data stores.

Check a DFD against these points: every process has at least one input and one output; flows go between a process and an entity or a store, never directly between two stores or two entities; each flow has a noun-phrase label; and Level 1 must be consistent with Level 0, keeping the same entities and the same data crossing the boundary. Recent HSC papers have asked students to distinguish a Level 0 and a Level 1 diagram and to use the correct symbols.

Infographics

An infographic combines short text, simple charts and visual cues to explain one idea quickly to a general audience, such as a grower checking status at a glance or a customer reading a report. Good ones apply the design principles from Year 11 and Year 12 data visualisation: one main message per panel, a clear reading order, a limited palette with enough contrast, labelled values with units and a source or date, and honest scales. An infographic simplifies, so it suits summarising the logic and results for non-technical readers, while a flowchart or DFD suits the team that will build or check the system.

Disruptive Effects of Intelligent Systems

🎯 (EC-12-06, EC-12-07)

📌 Investigate the disruptive effects of intelligent systems

• multitasking versus digital distraction
• working differently to complete the same task
• automation of enterprise and manufacturing processes
• impact of artificial intelligence (AI) on employment

A disruptive technology changes how work, services or industries operate, often faster than people and rules adapt. Intelligent systems are disruptive because they take over tasks that used to need human attention, and they also change how people spend their attention. Investigating the effects means weighing benefits and costs for different groups and reaching a balanced judgement.

Multitasking versus digital distraction

Intelligent devices let people do several things at once: reply to a message while a navigation assistant gives directions, or manage several dashboards at a time. That is multitasking, and it can improve productivity when tasks are simple or when a system handles routine parts. Digital distraction is the cost. Notifications and personalised feeds are designed by recommendation systems to capture attention, and research on attention suggests that switching between demanding tasks makes people slower and more error-prone than doing one at a time. In a workplace this can mean missed details, and in a car it is a safety risk. Enterprises manage it with notification settings, focus time and clear rules for device use, and designers can reduce it by showing only alerts that matter (which is itself a use of intelligent filtering).

Working differently to complete the same task

Intelligent systems often change how a task is done even when the outcome is the same. A farmer once walked the paddocks to check moisture and now reads a dashboard; a radiographer receives images with suspicious areas already highlighted; an accountant reviews the exceptions flagged by software instead of checking every transaction. The human role shifts from doing the routine work to supervising the system, handling exceptions and making judgement calls. That needs new skills such as interpreting data, questioning outputs and working with technology. It also allows more remote and flexible work, but can reduce face-to-face contact and hands-on knowledge of the underlying process.

Automation of enterprise and manufacturing processes

Automation uses systems to carry out processes without human intervention. In offices it covers tasks such as processing invoices, answering routine customer questions with chatbots and scheduling. In manufacturing it covers robots on assembly lines, sensor-driven quality checks and smart warehouses that move stock automatically. Benefits include consistency, speed, lower cost per item, round-the-clock operation and removal of people from dangerous tasks. Costs include the price of installing and maintaining systems, the risk that a fault stops the whole process, dependence on technology suppliers and the need to retrain staff.

Impact of artificial intelligence (AI) on employment

The employment effects are debated and varied, so a balanced answer names several. Tasks are automated: routine, repetitive work, whether manual or clerical, is the easiest to automate, so some roles shrink or change. New work appears: data analysts, sensor technicians, AI trainers and prompt writers did not exist a generation ago. Jobs change rather than vanish: many roles keep their title but the daily tasks change. Uneven effects: the impact differs by industry, location and skill level, and workers whose skills are not easily retrained may be affected most. Governments, employers and education systems respond with retraining, changed qualifications and debate about how the benefits of productivity gains are shared. Forecasts of how many jobs will be affected differ widely between studies, so treat any single number with caution.

Social and Ethical Issues

🎯 (EC-12-07)

📌 Investigate social and ethical issues to be considered when developing, implementing and using intelligent systems

An intelligent system makes or influences decisions that affect real people, so the developers, the enterprise that runs it and the people it affects all carry responsibilities. Investigating the issues means looking at them from each stage: developing, implementing and using a system. The issues below are the ones most likely to be examined, with an Australian setting.

Privacy and consent

Intelligent systems run on data, often personal data. The Privacy Act 1988 and the Australian Privacy Principles require organisations to collect only what is reasonably necessary, be open about what is collected and why, and keep it secure. People should give informed consent, not consent buried in long terms and conditions. The Privacy and Other Legislation Amendment Act 2024 has also created a statutory tort for serious invasions of privacy. Design questions include: is this data needed? Can it be de-identified? Can people opt out?

Bias and fairness

A system trained on biased or unrepresentative data reproduces the bias. A recruitment tool trained on past hires may favour the groups hired before, and a face recognition system trained mostly on some faces may work worse on others. Fair development means checking training data for gaps, testing outputs across groups and reviewing results regularly. The Australian Human Rights Commission and the anti-discrimination laws apply to automated decisions as well as to human ones.

Transparency, explainability and accountability

People affected by a decision should be able to find out that a system was used and how it reached its result. Rule-based expert systems can show the rules they fired. Neural networks are harder to explain, which is a problem when they decide who gets a loan or a benefit. The enterprise stays accountable even when software makes the recommendation: it cannot blame the algorithm. This is why the Robodebt scheme is studied in Australia. Its automated income averaging raised debts that were unlawful, and a Royal Commission examined the failures of oversight. Lessons included the need for human review, clear responsibility and a way for people to challenge the result.

Safety, reliability and human oversight

Systems that control physical equipment or influence health, money or liberty must be tested thoroughly and have a safe failure mode. Human oversight means a person can review, override or switch off the system, and knows when to. Overreliance is a risk in itself, since users may stop checking a system that is usually right.

Security

Connected devices and the data they hold are targets. Weak default passwords, unpatched firmware and unencrypted links have all been used to attack IoT devices, and an attacker who alters sensor data or model rules can change decisions. Enterprises apply the security measures from Year 11 and must notify the OAIC and affected people of eligible data breaches under the Notifiable Data Breaches scheme.

Employment, access and Indigenous data sovereignty

Automation can reduce some jobs, and people without access to devices, the internet or digital skills may be left behind by services that move to intelligent systems. Aboriginal and Torres Strait Islander Peoples have the right to govern data about their communities, Country and cultural knowledge. The CARE principles (collective benefit, authority to control, responsibility, ethics) and the Maiam nayri Wingara principles set out how such data should be collected and used; systems built on this data need community consultation and agreement.

Current and Emerging Technologies

🎯 (EC-12-06)

📌 Investigate current and emerging technologies associated with intelligent systems, including AI

Intelligent systems change quickly, so this dot point asks you to look at both established and emerging technologies and judge their potential and their risks. The technologies below are the ones most likely to appear in enterprises now. Emerging technologies are those still moving from research to everyday use, so check current information when you study, because details change quickly.

Machine learning and deep learning

Machine learning finds patterns in data to make predictions, and deep learning uses many-layered neural networks for tasks such as image and speech recognition. NESA's course specifications expect students to know that machine learning and statistical modelling can be used independently or together to analyse datasets and make predictions, without needing to understand the underlying logic. Enterprises use them for demand forecasting, fraud detection and quality inspection.

Generative AI

Generative AI creates new content, such as text, images, audio, code and video, by learning patterns from very large datasets. Large language models are a well-known example: given a prompt, they generate likely text one piece at a time. Enterprises use them for drafting documents, summarising meetings, customer support chatbots and writing code. The limits matter as much as the uses. Output can be fluent and wrong (sometimes called hallucination), reflect biases in the training data, reveal confidential information if staff paste it into a public tool, and raise copyright questions about training data and generated work. A responsible enterprise sets a policy on what may be entered, checks the output and keeps a person accountable. Students must also follow NESA's rules and their school's policy on the use of AI in assessment tasks.

Computer vision and speech

Systems that interpret images and sound power face unlock, quality inspection on production lines, medical image analysis, number plate recognition and live captions. They depend on quality training data and can fail in poor light or with unusual inputs.

Edge AI and IoT

Edge computing places processing close to where data is produced, such as inside a camera or a farm gateway, so decisions are fast and work without a constant internet link. It also reduces the data sent to the cloud. Small models running on microcontrollers make edge AI practical.

Robotics and autonomous systems

Robots and autonomous vehicles combine sensors, AI and actuators. Warehouse robots, autonomous haul trucks on mine sites, drones for crop monitoring and self-driving vehicles all make decisions in changing environments, which makes testing and safety the central issues.

Digital twins, AR and VR

A digital twin is a live virtual copy of a machine, building or process that is updated with sensor data and used to test changes and predict faults. AR and VR provide interfaces for training and remote inspection.

Blockchain and quantum computing

Blockchain records transactions in linked blocks that are hard to change, which supports supply-chain tracking of goods. Quantum computing is still experimental for most uses, but it may eventually affect optimisation problems and the encryption on which security relies.

Meeting Enterprise Needs

🎯 (EC-12-01, EC-12-06)

📌 Explain how intelligent systems combine innovative techniques and technologies to meet the needs of enterprise

Enterprises rarely use one technology alone. To meet a need such as cutting waste, serving customers faster or keeping people safe, intelligent systems combine several techniques and technologies so that each covers another's weakness. Explaining this means naming the need, showing the parts, and linking the way they fit to the result.

A pattern that repeats

1. Sense: IoT sensors, cameras, forms and databases collect data.
2. Communicate and store: networks, edge gateways and cloud storage move and keep it.
3. Decide: rules from an expert system apply known constraints, and machine learning provides predictions with probabilities; fuzzy logic or certainty factors handle uncertainty.
4. Act: actuators, automated workflows and messages carry out or recommend the response.
5. Learn and review: outputs are compared with what actually happened so rules and models are improved.

Worked example: a regional hospital reduces missed appointments

No single technique would do the job. The model finds patterns in past attendance that a person could not write as rules, but it cannot know the hospital's policy; the rules supply the policy and explain the actions; automation delivers the reminders at scale; and people review the results. Each element fills a gap in another.

Why combining works

• Strengths offset weaknesses: machine learning adapts but is hard to explain; rules are transparent but rigid.
• Data flows end to end: IoT, cloud and analytics share one stream instead of separate silos.
• Efficiency and consistency: automation applies the same decision every time, at any hour.
• Trade-offs remain: more parts mean more cost, more points of failure and more to secure and audit.

Infrastructure for an Intelligent Network

🎯 (EC-12-01, EC-12-08)

📌 Investigate the infrastructure requirements for an intelligent network in an enterprise with devices linked through an Internet of Things (IoT)

• networks with servers, local and cloud storage, and end-point devices
• communication links that enable the efficient control and flow of data

An intelligent network links many small devices so that data flows to where decisions are made and commands flow back. For an enterprise, the infrastructure has to carry large numbers of small messages reliably, keep the data safe, and keep working when part of the network fails. The Year 11 IoT and IoMe and IoT protocols sections give the background; this section applies it to an intelligent enterprise.

Networks with servers, local and cloud storage, and end-point devices

End-point devices are where data enters or actions leave the network: sensors (temperature, moisture, motion), cameras, scanners, controllers, actuators, and the phones, tablets and computers people use. Many are small and cheap, with little processing power, so they depend on the network behind them. Between the end points and the servers sit network devices: a gateway that speaks the devices' protocols, switches and access points that connect them, and a router with a firewall linking the enterprise to the internet.

Servers run the software. An application server hosts the expert system or decision support system, a database server holds the records, and a messaging service receives readings from devices. Servers may be on the premises or hired from a cloud provider.

Storage may be local (a network-attached storage device or an on-site server) or in the cloud. Enterprises usually combine both, and each has strengths.

Communication links that enable the efficient control and flow of data

The links have to match the device. Choose them by range, bandwidth (how much data), latency (how quickly), power use, reliability, security and cost.

Efficient control and flow of data also depends on how the links are used. An edge gateway filters readings and sends only what matters, so the internet link is not flooded. Lightweight messaging protocols such as MQTT let devices publish small messages to a broker that delivers them to subscribers. Quality of service settings give urgent traffic, like a fire alarm, priority over routine data. Devices belong on their own network segment, with default passwords changed and firmware kept updated, because a compromised camera should not lead to the finance server.

Data in an Intelligent IoT Network

🎯 (EC-12-02, EC-12-04)

📌 Explore collection, type, storage, processing, application and transmission of relevant and surplus data in an intelligent IoT network

An intelligent IoT network produces a constant stream of data, and much of it is not useful. Exploring the data means following it through six stages, collection, type, storage, processing, application and transmission, and separating relevant data (data that helps a decision or supports the purpose of the system) from surplus data (duplicated, out of range, unused or collected only "just in case").

The six stages

Relevant and surplus data

Suppose the probe reports every second. That is 86,400 readings a day. If the cold room only needs to know when the temperature moves outside a safe band, and to log a 15-minute average, then 96 averages a day, plus any alerts, carry the same decision-making value. The gateway has cut the transmitted readings by a factor of about 900. Most of the rest is surplus: near-identical readings repeated, values from a faulty sensor, or images of an empty room that no one will ever need.

Deciding what is surplus is a design choice, not just a technical one. The purpose has to be clear before data is collected. This links to data minimisation in the Australian Privacy Principles: collect personal information only when it is reasonably necessary for the organisation's functions, and destroy or de-identify it when it is no longer needed (APP 3 and APP 11). What looks like surplus today may become valuable for training a model, so enterprises balance the cost and risk of keeping data against its possible value, and record the decision. The Year 12 big data section discusses data at scale.

Simulation, Modelling and Automation

🎯 (EC-12-01, EC-12-05)

📌 Investigate the application of simulation, data modelling and the automation of systems in a range of enterprises

• education and training
• business analytics
• high-risk applications

A model, in NESA's glossary, is a representation that simplifies and explains the structure or workings of a system, and can be used to test and predict behaviour under limited conditions. Three related ideas build on models. Simulation runs a model over time to see how a system behaves without the real risk or cost. Data modelling represents data and relationships, and uses statistics or machine learning to describe or predict. Automation, in NESA's words, is transforming and manipulating data without user intervention. Together they let an enterprise test, learn and act at scale.

Education and training

Simulation lets learners practise in a safe, repeatable setting. A flight simulator trains pilots to handle an engine failure that no one would stage in a real aircraft. A nursing programme uses a manikin that responds to treatment, or a VR scenario that changes with the trainee's choices. In schools, science simulations let students change variables in an experiment that is too dangerous, slow or expensive to run. An intelligent tutoring system automates part of teaching by adjusting the difficulty to each learner's answers. The advantages are safety, repeatability, cost and immediate feedback. The limits are that a simulation is a simplification, and that some skills, such as dealing with a frightened patient, are still best learned in real settings too.

Business analytics

Enterprises model their data to understand what happened and to forecast what might happen. A retailer models sales against season, price and promotion to plan stock; a logistics firm simulates a change to delivery routes before rolling it out; an insurer models the likelihood of claims. Automation then acts on the results: reports are generated and emailed on schedule, reorders are triggered when forecasts fall short, and dashboards update themselves. This is where decision support systems and business intelligence meet. See the Data Visualisation page for how such results are shown. The limit is the quality of the model: a forecast built on out-of-date data or wrong assumptions gives confident but wrong advice.

High-risk applications

Where errors are costly or dangerous, simulation and modelling let organisations test before acting. Fire agencies model how a bushfire could spread across terrain in given weather to plan warnings and resources. Engineers simulate the response of a bridge or dam to extreme loads, and developers of autonomous vehicles test enormous numbers of scenarios in simulation before any road testing. Automation also removes people from danger: remotely operated or autonomous machines work in mines, chemical plants and disaster zones, and automated systems shut down equipment when a sensor detects a fault. The main risk is over-trust in the model. A simulation can only represent what its builders included, so validation against real data and a plan for failure are essential.

Intelligent Systems in Surveillance

🎯 (EC-12-03, EC-12-07)

📌 Explain the role of intelligent systems in surveillance

• closed circuit television (CCTV)
• biometric scanning
• customer loyalty schemes
• fraud prevention
• sniffing
• trolling

Surveillance is the monitoring of people, places or data. Intelligent systems have changed it by making monitoring continuous and automatic: cameras no longer need a person watching, and computers can search millions of records in seconds. That brings real benefits in safety and fraud prevention and serious concerns about privacy and fairness. The six named items differ in who is watched, why, and what the law expects.

Closed circuit television (CCTV)

CCTV cameras record video for a defined place. Computer vision adds intelligence: counting people, detecting a person in a restricted area, reading number plates or flagging abandoned bags, so staff review alerts instead of screens. It supports safety and crime investigation. Concerns include recording people who are not suspects, storing footage longer than needed and accuracy problems. In NSW, the Surveillance Devices Act 2007 regulates the use of optical and listening devices, and the Workplace Surveillance Act 2005 requires employers to give staff notice before using camera surveillance.

Biometric scanning

Biometric scanning identifies people by fingerprint, face or iris. Airports use automated gates that compare a traveller's face with their passport photo, and workplaces use them for building access. Matching is fast and hard to fake, but biometric data is sensitive information under the Privacy Act 1988 and cannot be reset if stolen. Bias in recognition accuracy across different groups and the risk of false matches also matter. The Office of the Australian Information Commissioner has published guidance on facial recognition and has investigated its use by large retailers. Consent, necessity and proportionality are the tests to apply.

Customer loyalty schemes

A loyalty card links purchases to an identified customer. Intelligent systems analyse the history to segment customers, predict what they will buy next and target offers. Customers get discounts, and the retailer gets detailed data on behaviour. The exchange is not always obvious: the data can reveal sensitive facts such as health conditions or family changes, and may be shared with partners. Australian Privacy Principles 5 and 6 require the retailer to tell members what is collected and why, and to use it only for the stated purpose or with consent. Direct marketing must let people opt out (APP 7), and the Spam Act 2003 sets the rules for commercial messages.

Fraud prevention

Banks and payment providers run machine learning models on every transaction, comparing it with the customer's usual pattern of spending, location and device. A purchase that looks unusual is flagged, blocked or queried by an SMS or app message. This protects customers and the institution and is one of the clearest benefits of surveillance. The costs are false positives, where a genuine purchase is declined while travelling, and the fact that everyone's transactions are monitored, not just those of suspected fraudsters. The model must also keep up with criminals, who change tactics.

Sniffing

Sniffing is capturing and inspecting the data packets that cross a network, using a packet analyser. Network administrators use it legitimately to troubleshoot and to detect intrusions, and intelligent security systems analyse traffic for unusual patterns. Attackers sniff unencrypted traffic, such as passwords sent over an unsecured Wi-Fi network, as in the evil twin attack covered in Principles of Cybersecurity. Encryption (HTTPS, VPNs, WPA3 Wi-Fi) protects the content. Using a sniffer to intercept other people's communications without authorisation can be an offence under Commonwealth laws such as the Telecommunications (Interception and Access) Act 1979 and the Criminal Code.

Trolling

Trolling is posting abusive or provocative messages online to upset people or disrupt a conversation. It is a kind of surveillance issue because platforms use intelligent systems to monitor huge volumes of posts, automatically detecting abuse and reporting or removing it, while trolls can use bots and fake accounts to amplify attacks. Automated moderation is fast but makes mistakes, missing context such as sarcasm or removing legitimate posts. In Australia the eSafety Commissioner can require the removal of seriously harmful content, and using a carriage service to menace, harass or cause offence is an offence under the Commonwealth Criminal Code.

A useful test is necessity and proportionality: is the monitoring needed for a clear purpose, is there a less intrusive way, and have people been told? Transparency, limited retention and human review of automated decisions reduce the harm.

AI and Efficiency in IoT

🎯 (EC-12-01, EC-12-06)

📌 Explain how AI supports efficiency in an IoT network

An IoT network can produce more data than people can watch and more decisions than they can make. Artificial intelligence supports efficiency by turning that stream into timely decisions and by reducing the resources needed to run the network. In NESA's glossary artificial intelligence is intelligence demonstrated by machines, and the Year 11 machine learning in IoT section introduced how models use IoT data.

Ways AI improves efficiency

• Predictive maintenance: a model learns the normal vibration and temperature of a pump. When the pattern drifts, it schedules a repair before a breakdown, so there are fewer unplanned stops and less wasted maintenance.
• Anomaly detection: the system flags readings that do not fit, whether a failing sensor, a leak or an attack, without a person watching every value.
• Edge filtering: a small model at the gateway decides which data matters, cutting bandwidth, storage and energy use.
• Energy and resource optimisation: smart buildings adjust heating, cooling and lighting to occupancy and weather, and irrigation systems water only where needed.
• Network management and security: AI can allocate bandwidth to urgent traffic and detect unusual device behaviour that suggests a compromised device.
• Automated control: combining sensors, models and actuators closes the loop with no person in it for routine decisions.

Worked example: predictive maintenance on a pump

1. A vibration sensor and temperature probe on a water pump send readings to the edge gateway.
2. The gateway compares each reading with the pump's normal range learned from past data.
3. A rising vibration pattern is flagged as abnormal, well before it stops the pump.
4. The gateway sends an alert and a summary to the cloud, and the maintenance system books a technician.
5. After the repair, the result is recorded so the model can be improved.

Measures of efficiency here would be hours of unplanned downtime, maintenance cost, and the amount of data sent to the cloud, each compared before and after installation.

Rules and Facts for an Expert System

🎯 (EC-12-08)

📌 Develop a set of rules and facts that could be used within an expert system to draw conclusions

Developing the knowledge for an expert system is a design task: decide what the system must conclude, list the facts it needs, and write rules that link them. This section builds the knowledge base for a Wi-Fi help desk that a school could give to students and staff, using facts and IF–THEN rules that draw conclusions.

Step 1: decide the goals and the facts

Start from the conclusions users need: the likely cause and what to do. Then list the facts the system can ask about, each written as a clear yes or no observation. Add intermediate conclusions where several facts point to one idea, because rules can then reuse them.

Step 2: write the rules

Write one idea per rule, with facts on the left and a single conclusion on the right. Use AND when all conditions must hold.

R1  IF all_devices_offline AND router_lights_off  THEN power_problem
R2  IF all_devices_offline AND router_light_red    THEN provider_problem
R3  IF one_device_offline                          THEN device_problem
R4  IF power_problem                               THEN check_router_power
R5  IF provider_problem                            THEN restart_router_call_provider
R6  IF device_problem AND airplane_mode_on         THEN turn_off_flight_mode
R7  IF device_problem AND password_changed         THEN re_enter_password

The chain works as in the expert-system demo. Given all_devices_offline and router_lights_off, forward chaining fires R1 (adding power_problem) and then R4, giving the advice to check the router's power.

Step 3: check and test the rules

• Conflicts: can the same facts lead to two different conclusions? If so, make the rules more specific or add a priority.
• Gaps: is there a plausible combination of facts with no rule? Here, no rule handles a single offline device with neither flight mode nor a password change, so add a rule such as R8 IF device_problem THEN restart_the_device, which gives a sensible extra step in every single-device case.
• Redundancy and loops: remove duplicate rules, and make sure no rule chain leads back to its own starting fact.
• Vocabulary: use one name for one idea, or the engine will not match facts to rules.

The last test shows why testing matters: it exposes a case the rules do not cover. In the demo above, edit or add rules and check that the trace still reaches the advice you expect.

Certainty Factors and Decision Trees

🎯 (EC-12-05, EC-12-08)

📌 Apply certainty factors to construct a decision tree for a proposed expert system

Real evidence is rarely certain: a leaf looks "a bit yellow", or a sensor is sometimes wrong. A certainty factor (CF) is a number that shows how strongly a piece of evidence or a rule is believed, from 0 (no confidence) to 1 (complete confidence). Many systems also use negative values to show belief against a conclusion, but this section uses 0 to 1. Applying certainty factors to a decision tree means giving each conclusion a CF, so the system reports how sure it is as well as what it thinks.

The arithmetic

1. One rule: CF of the conclusion = CF(rule) × CF(evidence).
2. Several conditions joined by AND: the evidence is only as strong as the weakest condition, so take the minimum of their CFs. With OR, take the maximum. (This is the usual convention; your teacher may set one, so state the one you use.)
3. Two rules supporting the same conclusion: combine their CFs with CF1 + CF2 × (1 − CF1). The result is higher than either but never above 1.

Worked example. Rule A says IF leaves_yellow AND soil_wet THEN overwatered with rule CF 0.8. The user is 0.9 sure the leaves are yellow (seen by eye) and 0.7 sure the soil is wet. Rule B says IF no_drainage THEN overwatered with rule CF 0.5, and the user is completely sure (1.0) the pot has no holes.

1. Evidence for rule A = min(0.9, 0.7) = 0.7.
2. CF from rule A = 0.8 × 0.7 = 0.56.
3. CF from rule B = 0.5 × 1.0 = 0.50.
4. Combined = 0.56 + 0.50 × (1 − 0.56) = 0.56 + 0.50 × 0.44 = 0.56 + 0.22 = 0.78.

Two moderate pieces of independent support give 0.78, more than either alone. Try other values in the calculator.

Constructing the decision tree

NESA describes a decision tree as representing all possible combinations of decisions and their resulting actions, with branches leading to another decision or a final action. Follow these steps for a proposed plant-care expert system.

1. List the final actions the system might recommend: repot and reduce watering, reduce watering, water thoroughly, add fertiliser, no action.
2. Choose the questions that split the cases, starting with the most useful one (are the leaves yellow?). Every question needs labelled Yes and No branches.
3. Assign a rule CF to each path from an expert's confidence or from past records of how often that path proved right. Here a blocked drainage path is 0.9 because the combination is almost always a watering problem, while adding fertiliser is only 0.5 because several other causes give the same symptom.
4. Draw the tree in NESA notation: rectangles for questions and actions, straight labelled branches, and each path ending in a final action, with its CF written beside it.
5. Set a threshold. For example, if the final CF is below 0.6, the system asks another question or refers the user to a person.

Using the tree. For the first action, take the minimum of the evidence CFs: min(0.8, 1.0, 0.7) = 0.7. Multiply by the rule CF: 0.9 × 0.7 = 0.63. Since 0.63 is above the 0.6 threshold, the system recommends repotting and reducing watering, and reports its certainty as 0.63. If the user had been only 0.6 sure about the drainage, the result would be 0.9 × 0.6 = 0.54, so the system would ask the user to check the pot before advising.

Verifying Data Sources

🎯 (EC-12-03, EC-12-10)

📌 Verify the sources of data used in a decision support system

A decision support system is only as reliable as the data behind it. Verifying a source means checking, with evidence, that the data is what it claims to be and is fit for the decision. Do it before the data is loaded, and repeat it when the source is updated. The steps below give a method a student could carry out and record.

Criteria for a trustworthy source

• Authority and provenance: who collected the data, how and why? Government and research bodies (for example the Bureau of Meteorology or the Australian Bureau of Statistics) are generally more reliable than an unnamed website.
• Accuracy: does the data agree with a second independent source, and were the instruments calibrated?
• Currency: is it recent enough for this decision, and does it have a date?
• Completeness: are there gaps, missing periods or missing groups?
• Consistency and validity: do the values follow the expected type, format and range, and do totals reconcile?
• Relevance and bias: does it match the question, and could the way it was collected favour one outcome?
• Permission: is there a licence to use it, and does it include personal information subject to the Privacy Act 1988?

Methods for checking

1. Identify every data source that feeds the DSS and record its owner.
2. Read the documentation or metadata: date, method, units and known limits.
3. Cross-check a sample against a second source or against a known figure.
4. Run validation checks: data type, allowed range, format, required fields and duplicates. See Data Science: reliability for more on data quality.
5. Test sensors against a reference reading, for example a moisture probe against a lab measurement.
6. Record the result in a verification log and decide to accept, correct or reject the source.

Building a Knowledge Base

🎯 (EC-12-08, EC-12-11)

📌 Use a flowchart to develop a knowledge base of IF-THEN rules to be used by an expert system

A flowchart is a good starting point for a knowledge base because experts often describe their reasoning as a sequence of questions. NESA's flowchart notation makes that sequence visible, and each complete path through the chart can then be turned into one IF–THEN rule. This section develops the knowledge base for a printer help desk in four steps.

Step 1: draw the flowchart

Capture the expert's questions in order, using the four NESA symbols. The chart below records how a technician finds a printer fault.

Step 2: turn each path into a rule

Read from START to each action. Every decision on the path becomes a condition: a Yes branch means the condition is true, a No branch means it is not. Join the conditions with AND, and the action becomes the THEN part.

Step 3: tidy and test the rules

• Simplify: NOT conditions can be replaced by a fact worded the other way round, such as printer_offline, so a user answers simply yes or no.
• Check coverage: the four rules cover every combination the chart allows, because a chart with labelled Yes and No exits cannot leave a case out.
• Test: feed each path in as facts and confirm the expected action. Add a test that should give no answer, such as no facts supplied, to see how the system behaves.

Step 4: enter the rules into the expert system

Type the rules into expert system software, which, according to NESA's course specifications, should let you develop IF–THEN rules, add, remove and edit them, and display the rules used to reach a conclusion in text and as a decision tree. Run each test case, then read the rules-used display and compare it with your flowchart path. They should match. Changes are then made by editing a rule, not by rewriting the program: if a new printer model adds a "toner low" light, add one rule and one branch of the chart.

The process also runs in reverse. Writing the rules first and drawing them as a flowchart is a good way to spot missing questions or unreachable actions.

Designing an Automated Smart System

🎯 (EC-12-08, EC-12-11)

📌 Design and model an automated smart system using a range of inputs and outputs

An automated smart system senses its environment, decides using rules or a model, and acts with little human involvement. Designing and modelling one means working out what goes in, what comes out and what happens in between, then representing the design with diagrams and testing it before anything is built. The example is a smart greenhouse for a plant nursery, using the notation from the Course Specifications.

Step 1: define the need and the boundary

Seedlings die if the greenhouse gets too hot or too dry. The nursery wants the greenhouse to stay within a comfortable range overnight and on weekends without anyone attending, and to alert the grower if something goes wrong. Manual override and safety limits are requirements, not extras.

Step 2: choose the inputs and outputs

Step 3: model the decision logic

NESA's specification shows a decision tree for the comfort control of a smart house. The same form suits any rule set with several inputs and several outputs: a column for each input and each output, and one row for every combination.

Step 4: model the whole system

A system flowchart shows the main processes and devices, and the direction of data between them. In NESA's notation, online input symbols are the sensors, online display is the dashboard, storage holds the rules and the log, and the telecommunications link and cloud show the connection to remote analytics.

Step 5: test the design before building

Model it with simulated data first. In a spreadsheet, a block-based coding tool or a simulator, feed in made-up readings and check that the outputs match the design, including the boundaries.

Think about failure as well as normal running: what happens if the pump sticks on, the internet drops or the sensor gives an impossible value? A good design uses limits (a maximum pump time), a manual override and a fall-back that keeps plants alive.

Implementing Automated Processing

🎯 (EC-12-08)

📌 Implement automated processing using software

Implementing automated processing means turning the design into software that reads data, applies the rules and produces outputs without anyone pressing a button each time. The Course Specifications do not fix a language, so use the tools available: a spreadsheet, a text-based language such as Python, a block-based tool, or expert system software. The same greenhouse design is implemented below in two of them.

Implementation 1: a program

The function decide holds the rules, and the loop feeds in a series of simulated readings, standing in for real sensors.

MOISTURE_MIN = 40      # percent: water below this
HOT = 30               # degrees Celsius: fan high above this
WARM = 20              # degrees Celsius: fan low from this up

def decide(moisture, temperature, rain_forecast):
    """Return the actions for one set of readings."""
    actions = {"pump": "off", "fan": "off"}
    if moisture < MOISTURE_MIN and not rain_forecast:
        actions["pump"] = "on"
    if temperature > HOT:
        actions["fan"] = "high"
    elif temperature >= WARM:
        actions["fan"] = "low"
    return actions

# Simulated readings: (moisture %, temperature, rain forecast)
readings = [(38, 24, False), (55, 32, False), (35, 28, True), (30, 18, False)]

for moisture, temperature, rain in readings:
    result = decide(moisture, temperature, rain)
    print(f"moisture {moisture}%, {temperature} C, rain {rain}: "
          f"pump {result['pump']}, fan {result['fan']}")

Running it produces:

moisture 38%, 24 C, rain False: pump on, fan low
moisture 55%, 32 C, rain False: pump off, fan high
moisture 35%, 28 C, rain True: pump off, fan low
moisture 30%, 18 C, rain False: pump on, fan off

The rules are in one place, so a threshold changes in a single line. To use real sensors, replace the list with a function that reads them, and replace print with code that switches a relay. Add a loop that waits between readings, and a limit on pump run time.

Implementation 2: a spreadsheet

Spreadsheet formulas recalculate automatically whenever the input cells change, which is NESA's own example of automation. Put readings in columns B to D, and decisions in E and F. Only the IF function and relational operators are needed.

Fill the formulas down for each new row of readings. Add conditional formatting so an ON pump cell turns green and a HIGH fan cell turns orange, and a chart of moisture over time. A macro can be recorded to import the latest readings and sort them, so the whole sequence runs with one click. Both methods implement the same design, and differences between their results reveal mistakes.

Testing the implementation

Use the methods named in the Course Specifications. Functional testing checks each rule against expected results. Simulated data is made-up data that includes boundary and extreme values, so it is safe to try before real sensors are connected. Live data uses the real readings once it is safe to do so. Acceptance testing lets the client confirm that the system meets their needs.

Assessing Decision Support Output

🎯 (EC-12-05, EC-12-10)

📌 Assess the output produced by a decision support system, including graphing to detail the success of decisions, and comparing the proposed versus the actual outputs

Assessing a decision support system's output means judging how good its recommendations were. The clearest evidence comes from comparing what the system proposed with what actually happened, then graphing and measuring the difference. The example below uses a supermarket DSS that proposes a weekly order for one product, and the sales that followed (sample data).

Step 1: define what success means before you look

Choose measures and a tolerance in advance so the assessment is fair. Here, a week counts as a success if the proposed quantity is within 10% of actual sales. Other criteria can be used as well: accuracy (how close), timeliness (was the advice ready when needed), consistency (similar inputs give similar advice), usability (did staff understand and use it) and value (did stock-outs or waste fall).

Step 2: tabulate proposed against actual

In a spreadsheet, with proposed in B and actual in C: error =C2-B2, absolute error =ABS(C2-B2), percentage error =ABS(C2-B2)/C2, the test =IF(F2<=0.1,"Yes","No"), and the summary figures with SUM, COUNT and the mean function.

Step 3: graph it

A line graph with both series on one set of axes shows where the two diverge, which a table hides. Ring the weeks outside tolerance. A scatter graph of proposed against actual, where perfect advice falls on a straight line, is another good choice, and a column chart of the errors shows their size and direction.

Step 4: calculate and interpret

• Mean absolute error: the absolute errors add to 75, so 75 ÷ 8 = 9.4 units per week on average.
• Mean percentage error: about 6.3%, so on average the proposal is within 7% of actual sales.
• Success rate: 6 of 8 weeks (75%) were within 10%.
• Bias: the signed errors add to +1, so there is no consistent tendency to over-order or under-order. Total proposed (1,195) and total actual (1,196) are almost the same, which hides the week-to-week misses.

The pattern in the misses matters more than the average. Weeks 4 and 7 are the largest errors, and they go in opposite directions. Investigate their causes: was there a promotion, a public holiday, bad weather or a data fault? If a cause can be found and fed into the model as a new input, the DSS can be improved. If not, the manager should treat the recommendation as a guide and keep a safety margin.

Step 5: judge and recommend

A balanced judgement for this sample: the DSS is reasonably accurate and unbiased, and it is useful for planning, but it is not reliable in unusual weeks, so the manager should review weeks with promotions or events. Recommend adding the promotion calendar as an input, continue comparing proposed and actual each week, and retest after changes. Eight weeks of data is a small sample, so the conclusions should be confirmed with more.

Expert Systems and Efficiency

🎯 (EC-12-01, EC-12-06)

📌 Explain how expert systems contribute to the efficiency of intelligent systems

• supercomputers
• digital assistants
• autonomous vehicles
• streaming services

Expert systems contribute to the efficiency of larger intelligent systems by supplying fast, consistent decision rules that other components, such as machine learning, powerful hardware and sensors, cannot supply on their own. Their contribution is usually the rules: policy, safety limits, routing and expert know-how. Four examples show this in different settings.

Supercomputers

Supercomputers run enormous calculations in parallel, for weather forecasting, climate models, engineering simulation and scientific research. Efficiency comes from using that expensive capacity well. Rule-based scheduling systems decide which jobs run where and when, and expert knowledge about a model sets its parameters and checks results against known physical limits, so the machine does not waste hours on impossible runs. The forecasts that result (for example the weather warnings that farms and emergency services use) show the value: better decisions, made earlier.

Digital assistants

A digital assistant, in an app, a phone or a call centre, uses rules and models together. Rules map recognised requests to actions ("IF the intent is set alarm THEN open the clock function") and enforce policy such as identity checks. Machine learning recognises speech and interprets varied wording. Efficiency shows as time saved: routine questions are answered instantly at any hour, and only harder ones go to a person, who receives the details already collected. The limits are wrong answers and misheard requests, so a route to a human is essential.

Autonomous vehicles

An autonomous vehicle combines sensors (cameras, radar, laser scanners, GPS) with machine learning that recognises objects and rules that encode traffic law and safety limits, such as stopping distances and speed limits. The rule layer is easy to inspect and certify, and the learning layer copes with messy real-world sights. On mine sites such as those in Western Australia's Pilbara, autonomous haul trucks follow planned routes around the clock, which supports steady operation and removes people from hazardous work. Efficiency is measured in things like fuel use, running hours and incident rates. Safety testing and responsibility for accidents remain the major challenges.

Streaming services

A streaming service recommends what to watch from your history and from the habits of similar viewers, a form of machine learning. Rules then apply constraints: licensing by country, age ratings and parental controls, and business rules on what to promote. Other intelligent components improve efficiency behind the scenes, for example adjusting video quality to the available bandwidth and caching popular titles closer to viewers to reduce network load. The results are less searching for the viewer and more efficient use of the service's network, though recommendations can narrow what people see and depend heavily on collecting viewing data.