🤖 AI modellaréna

Hierarchical matchup expectations (lane → champion → matchup) with shrinkage. Judges 'going even into a counter' fairly.

+ Interpretable, tiny (JSON), robust to sparse data

− A baseline, not a predictor

Linear model over 15 matchup features, stacked on the rating prediction.

+ 15 params, trains in <1s, well-calibrated, easy online updates

− No feature interactions

Two hidden layers (48·24) learning non-linear boundaries.

+ Expressive, still browser-trainable

− Overfits small data

200 histogram-boosted trees — the XGBoost/LightGBM family, king of tabular data.

+ Thresholds & non-linearity, scale-free

− Larger model, awkward online updates

Pairwise feature interactions via latent vectors (k=8).

+ Interactions with few params

− Needs gradient clipping

Models P(x|win) and P(x|loss), inverted via Bayes. Temperature-calibrated.

+ Single-pass training, ultra-light

− Overconfident on correlated features

Samples thousands of virtual games from player performance distributions.

+ Uncertainty & close-game rate, works with zero training data

− Simple generative assumptions

Validation-weighted logit pooling of all models, with disagreement detection.

+ Consistently beats the best single model

− Needs the full pack

Chrome built-in AI (Gemini Nano) → WebLLM → Transformers.js → template fallback turns the numbers into a natural-language read.

+ Zero server traffic, adapts to your device

− Heavy engines need a one-time download