Review

Core workflow

configuration -> data -> computation -> plot -> summary

Experiment as a script

Use a main-like workflow so the experiment reads in order.

Keep configuration visible:

config = ExperimentConfig(seed=4, points=8, true_slope=2.0, noise=0.2)

Inputs, outputs, and checks

Name inputs and outputs before trusting a result.

Useful checks:

Noisy linear data

The experiment uses:

y = slope * x + noise

Synthetic data lets you know the hidden truth.

Fitting by search

For each candidate slope:

Choose the slope with the lowest loss.

This chooses only among the supplied candidates, with the intercept fixed at zero.

Plotting

Plot both:

• noisy data points;
• fitted line.

Then write what the plot shows.

Saving a summary

Save enough to understand or rerun the experiment:

• seed;
• points;
• true slope;
• fitted slope;
• loss;
• observation.

Reading code as a report

A reader should be able to answer:

• What was fixed?
• What was varied?
• What was measured?
• What did the plot show?
• What result was recorded?