Underfitting and OverfittingVisualization

Here is the dartboard with real models. The dashed curve is the true harvest pattern of the estate, which nobody knows exactly. Neha has records from five years; each year's 12 readings scatter around the truth because of weather and luck. Fit the same kind of curve to each year separately and compare the five fitted curves and their average. The chart under the readouts keeps bias², variance and their sum for every degree you try. Move the degree and complete the missions.

Degree of the curve fitted to each year: ${degree}

Bias² measures how far the average curve is from the truth; variance measures how much the five curves spread around their average. A simple curve has high bias and low variance; a bendy one has low bias and high variance. The expected error of a model on new data is

The noise — the luck of each year — no model can remove. The best model makes the sum of the other two smallest: here degree 316, the bottom of the red line. It is the U-shaped validation curve you met in Train, Test and Validation, seen from the inside.