Choose the conditions
Start with all three ingredients present, or experiment by clicking the input circles directly on the map. A circle switches between present (1) and missing (0).
Welcome backYour network and lesson progress are saved on this device.
Interactive lesson · The Fire Triangle
Teach a Neural Network What Fire Needs
Heat, fuel, and oxygen must all be present for fire. Train this small network to discover that relationship one example at a time.
No machine-learning background needed. Success means watching a wrong guess improve after you teach the correct rule.
Live network map
See the lesson move through the network
3 inputs · direct connection · 1 output+ Positive weight − Negative weight
Preparing0% checks passing
Preparing exampleFire Triangle auto-training
Inputs—
Prediction—
Correct target—
Error—
The trainer shows each visible round so you can watch predictions, errors, and weights change.
Tap a condition above or on the map to switch it between present and missing.
Try it: click or tap an input circle to switch that condition between present (1) and missing (0). Thicker lines have more influence.
Math behind this step
Waiting for a prediction
Start hereAdd each condition’s vote, then turn the total into an output confidence.
See the technical equation
z = Σ(wᵢxᵢ) + bŷ = σ(z)Choose conditions and make a prediction to see the network’s calculation in everyday language.
The simple idea
Each condition gets a vote. The network adds those votes to its starting opinion, then converts the total into a percentage.
Guided experiment
Train the network by giving it examples
A neural network learns from correction, not from being handed the rule. Complete the four steps, then repeat them with different fire conditions.
Next action
Choose a Fire Triangle scenario to begin.
0 training examples completed
🔥Fire0 examples
○No oxygen0 examples
○No fuel0 examples
○No heat0 examples
★0 cyclesBalanced training
Review completed detector tests
Level 2 unlocked
Make the network a little more capable
The simple network is now reliable on all four Fire Triangle cases. Ready to add suppression? You will add one new concept at a time—an input, then a hidden layer, then a “Safe” output—and see each change appear on the map.
Your upgraded network is ready. Choose a new example:
Level 3 unlocked
Build a larger fire-spread network
The current network can decide Fire and Safe. Now expand it to consider Wind, process the information through a second hidden layer, and predict a third answer: Spread Risk.
The expanded network is ready. Choose a five-input example:
Teach the network when the landscape is ready to burn
Add high temperature, low humidity, and dry vegetation. These conditions do not prove a fire exists; together they describe elevated wildfire risk.
The examples from here forward are synthetic rules created for this lesson, not field wildfire data.
Choose an environmental example:
Combine evidence from wildfire sensors
Add smoke, infrared heat, and poor air quality. One reading can be misleading; several agreeing sensors create a stronger detection alert.
These sensor readings are simulated so you can see how several signals can be combined.
Choose a sensor example:
Teach the detector when evidence is trustworthy
Real sensors can fail or report noisy readings. Add Sensor Health and train a Reliable Alert output that requires both evidence and trustworthy equipment.
A real reliability model would require calibrated sensors and verified failure data.
Choose a reliability example:
Check examples the network did not train on
A model is useful only if it can handle new situations. This evaluation freezes the weights and scores a separate, balanced synthetic test set.
No test has run yet.
Choose when the detector should raise an alert
A lower Reliable Alert threshold catches more possible fires but can create more false alarms. A higher threshold is quieter but may miss danger.
Use the system you trained
Set conditions by clicking inputs on the map, then ask the completed detector to analyze the current environment.
Your trained detector is ready for an analysis.
Educational simulation only. A real emergency system requires validated field data, calibrated sensors, monitoring, and human oversight.
Ask for a prediction
The network combines the conditions currently shown on the map and returns a confidence. Try one prediction, toggle a map input, then predict again to compare.
Network confidence
-
Choose conditions, then predict.
Tell it the correct answer
First choose conditions and make a prediction. Then the Fire Triangle will provide the correct answer.
Correct answerWaiting…
Manual target controls
Watch what changed
The network measures how far off it was, traces the mistake backward, and adjusts each useful connection. Now toggle a map input and predict again to test the new behavior.
Before—→After—
Train an example to see the confidence move.
Waiting for your first prediction
Choose a fire scenario above to begin.
Inputs
Edit input names, values, and correction strength manually
Adjust the conditions manually. For this lesson, keep each value between 0 and 1.
Correction strength controls how boldly the network changes after a mistake. The default 0.3 is easy to see while remaining stable.
Optional lab: change the network yourself
The guided Fire Triangle lesson works without these controls. Open them when you want to experiment beyond the walkthrough.
Architecture
What the network is changing
Importance (weight)
Each connection stores how strongly one value should support or oppose the next neuron or output. Training changes that importance after mistakes.
Starting opinion (bias)
This is the neuron's tendency before any condition votes are added. Training adjusts it too.
Confidence converter (activation)
Sigmoid turns each combined score into a 0–100% signal. Only the final layer produces the named output confidences.
Tracing the mistake backward (backpropagation)
The network compares every output with its correct answer, assigns responsibility through the hidden layers, then adjusts from the outputs back toward the inputs.
Open technical inspector
Optional: inspect or edit every number after you understand the guided training loop.
Weights and biases
These values update after every training example.