🧠 Supervised, Unsupervised and Reinforcement Learning

About this simulation

Three glowing platforms float side by side, each running the same raw data through a different learning paradigm. The left platform is fed color-coded labels and learns a decision boundary that separates them. The middle platform never sees any labels — it only sees positions, and groups points into clusters by discovering structure on its own. The right platform is a single agent hopping across a reward grid, learning by trial and error which path earns the highest score.

How it works

• Supervised — a small neural boundary surface is nudged epoch by epoch to minimize how many labeled points fall on the wrong side, exactly like training a classifier on labeled examples.
• Unsupervised — k-means centroids drift toward the mean of their nearest points every epoch, with no labels available; only the geometry of the data guides the grouping.
• Reinforcement — an agent explores a grid of reward tiles, updates a simple value estimate after every hop, and gradually shifts from random exploration to exploiting the best-known path.
• All three run on the very same underlying point cloud, so you can directly compare how differently each paradigm treats identical data.

Controls

• Training speed — how many epochs per second each learner advances.
• Label noise — percentage of supervised labels that are deliberately wrong, showing how mislabeled data hurts convergence.
• Reward sparsity — fraction of the RL grid's reward tiles that are hidden/zeroed, simulating sparse-reward environments.
• Dataset shape — switch between blobs, an interleaved spiral, and concentric rings to see how geometry affects each method.
• ↺ New data — resample the underlying point cloud and reset every learner. Drag to rotate, scroll to zoom.

Did you know?

Supervised learning needs the most human effort (labeling), unsupervised learning needs the least but gives the least control over the output, and reinforcement learning needs neither labels nor pre-collected data — only a way to try actions and observe rewards, which is why it is used for game-playing agents and robotics.