MLArenaa PLARENA lab

machine learning, playable — your clicks are the training set

Linear / Polynomial Regression

Fit a curve through data points, then query it — including the future. Stars, residuals, readable weights, the Forecast scene, and the degree dial.

continuous outputs · closed form · Fit view · Hand fit

Forecast (regression in time)

Your food truck's daily sales: fit a curve to 24 past days, then read next week off it. The degree dial decides whether the forecast is honest — or a hallucination.

continuous target · time axis · the future = extrapolation · Reveal

Logistic Regression

From scores to probabilities: one S-curve, a straight boundary, calibrated confidence. Watch gradient descent raise the boundary from nothing.

like/dislike · sigmoid view · watch training · Hand fit

MLP (Neural Network)

A committee of straight lines that votes itself into a curve. Slide hidden units from 0 (it is logistic) to 16 and watch capacity grow.

hidden-units dial · anatomy overlay · watch training

Compare & Bet

All three lenses on one canvas, one dataset. Place your bet, run the pinned scenarios — XOR upset, the overfitting Scissors — and watch the champion rotate.

POE bets · scenarios · A/B slots · champion rotation
1 · PERCEPTS — sensors: dish features + the diner's ratings, collected by the house
Menu — taste something, rate it. The model learns you.
3 · ACTIONS — actuator: recommendations, served back to the diner
Recommendations — the 5 unrated dishes where f predicts the HIGHEST score (ranked by ŷ, not by error) · blue rank rings mark them on the map
Environment — what kind of world is this? (live)
observablePartially
deterministic—
agentsSingle
episodic / sequentialEpisodic
static / dynamicStatic
discrete / continuous—
Each dish is judged on its own — no action compounds into the next state. (Search & Game Arena are the sequential counterexamples.) Flip the persona or scenario and watch the deterministic row change. Learning-agent anatomy: performance element = the fitted f ranking dishes · learning element = the fit routine · critic = the Metrics panel · problem generator = you (invented dishes, the degree dial, the scenarios). The house runs the agent; the diner's taste is what f estimates.
2 · AGENT FUNCTION — f: (spiciness, price) → rating. The learnable middle — most of the agent lives here
Controls
Data
Target seed
Lens degree 1 ?
Scenarios:
Place your bet: which lens will score best on the unrated items?
Taste space — x: spiciness · y: price · grey axis bars: dish counts per bin (the data's distribution — hover them) · fit updates as you rate
hover the space or the menu to inspect an item
Metrics — counts, not percentages
baseline (majority)—
train—
held-out pool—
Loss — train (solid) vs held-out (dashed): watch the scissors

Work in progress

Plarena is under active development — rough edges, bugs and placeholder content are all expected. Found something broken, confusing or missing? Have an idea? Please tell Xian Su at feedback@plarena.app. Every piece of feedback makes the next version better. Thank you.