Experiment 6 · about 12 min

Throw and predict

0/4 tasks

The robot has a throwing arm whose power goes from 1 to 10. It doesn’t know where the ball will land for each power. Make some throws, collect data and train a line; then let the model predict the power it takes to land in the basket.

Pick a power and throw. Wherever the ball lands, a point is added to the chart.

1. Collect data

0 throws

Throw with different powers. Every throw is a data point made of a power and a distance.

Power5

3. Aim for the basket

0 in a row

The basket is 3.5 m away. You don’t pick the power: the model works it out from the line.

First make at least two throws and find the line.

2. Find the line

Move the line closer to the points with the sliders. The red lines show how far each point is from the line: that is the error. Then watch the model do the same job.

Slope: how much further for 1 more power0.4Start1 m

The line as Python code

power = 5
distance = 0.4 * power + 1
# ≈ 3 m
target = 3.5
power = (target - 1) / 0.4

Tasks

Make at least 6 throws (waiting)
Set the line yourself with the sliders: on average closer than 0.35 m to the points (waiting)
Land 3 baskets in a row with the power the model suggests (waiting)
Land the far basket at 9 m (waiting)

What’s going on?

This is regression: the model doesn’t pick a group, it predicts a number. Here it predicts distance from power, and turned around, power from distance.

The model takes how far each point is above or below the line, squares it, adds them all up and takes a step that makes the total smaller. This method is called least squares.

The model can only be trusted where it has seen data. A real throw isn’t a straight line: the harder you throw, the more the air slows the ball down. A line trained on gentle throws makes the ball land short at the far basket.

Try this: throw only with powers from 1 to 5, train the model and aim for the far basket. Then add a few strong throws, train the model again and try once more.

Related lessons