Bias-variance trade-off explained as a G検定 staple

The explainer frames bias and variance as the two components behind a model's prediction error, then states the decomposition as バイアス² + バリアンス + 既約誤差, separating error you can control from irreducible noise. From there it lists levers: complex models, richer feature engineering and hyperparameter tuning to cut bias; more training data, stronger regularization and ensemble learning to cut variance.

The central point is the trade-off: as model complexity rises, bias falls but variance rises. For practice, the explainer suggests drawing training curves—a wide gap between training and validation error signals high variance (overfitting), while both being high signals high bias (underfitting)—and prescribes opposite fixes for each.

The piece is written partly as exam prep, noting bias and variance appear often on the G検定 syllabus and pointing readers to a simpler kids' version and to study resources, including the official JDLA-supervised textbook.