One constant laboratory
The same grid, obstacles, goal and controls keep every comparison fair.
SupraLab AI
Interactive learning module
From randomness to a network that learns.
One robot, one laboratory, six ways to decide. Inspect every calculation, train the models and watch genuine learning emerge.
- Real-time simulation
- Visible calculations
- French and English
The principle
Compare what is programmed with what is learned.
The environment never changes. Only the way the robot chooses its next move evolves.
One robot
Visitors follow the same agent and clearly distinguish an action, a strategy and learning.
Every calculation visible
Rules, Q-values, weights, activations, errors and gradients are explained as they act.
Learning progression
Six levels that build intuition.
Every level answers the limits of the previous one and introduces a new family of decisions.
- 01Does not learn
Random agent
A minimal baseline that separates occasional success from progress.
- What the robot uses
- A randomly chosen direction
- Mechanism
- Random sampling
- 02Does not learn
Symbolic AI
A coherent strategy written entirely by the developer.
- What the robot uses
- Position, goal and rules
- Mechanism
- Explicit conditions
- 03Learns
Q-learning
The robot improves decisions through trials, rewards and penalties.
- What the robot uses
- State, actions and rewards
- Mechanism
- Q-table updates
- 04Learns
Perceptron
A first decision boundary built from numerical inputs.
- What the robot uses
- Inputs and weights
- Mechanism
- Supervised learning
- 05Fixed weights
Multilayer network
A hidden layer demonstrates forward propagation and intermediate representations.
- What the robot uses
- Inputs, weights and activations
- Mechanism
- Forward propagation
- 06Learns
Backpropagation
Loss produces gradients that progressively correct the network’s weights.
- What the robot uses
- Error, gradients and examples
- Mechanism
- Gradient descent
Learn by manipulating
The visitor is not watching a black box.
Every experiment connects the robot’s visible behaviour to the numbers that produced its decision.
Simulate
Move step by step or automatically, at four speeds.
Train
Accelerate Q-learning and train the supervised models.
Inspect
Read inputs, weights, probabilities, errors, gradients and the loss curve.
Compare
Observe what remains static and what genuinely improves through experience.
Actual scope
A readable experiment, not a production model.
SupraLab AI prioritises understanding the mechanisms. This simplicity is a deliberate educational choice.
- A fixed map keeps all six approaches comparable
- Small models remain fully inspectable
- Supervised data is generated from accessible laboratory positions
SupraLab AI
Follow the module launch.
Receive the interactive experience link when it opens in the SupraLearning ecosystem.