AI and programming classes in Roundhay

The learner downloads four Census 2021 tables from the Nomis API for the 2,607 output areas of Leeds, which hold 341,381 households, 98,083 of them with no car or van. A gradient boosting model predicts each area's no-car share from its population density, its share of one-person households and its share of flats. Two settings are searched: the learning rate, anywhere from 0.001 to 1, and the subsample, the fraction of areas each tree sees, from 0.3 to 1. Every candidate is scored by cross-validation on 70% of the areas; the other 30% are kept back for a final check. Guessing the average for every area is wrong by 15.62 points.

The grid reveals why it loses. Reading across its rows, the error is about 14.5 at a learning rate of 0.001, about 9.6 at 0.01, about 8.6 at 0.1, and between 11.63 and 27.24 at 1.0. The learning rate decides almost everything, and the subsample hardly matters except when the rate is far too high. Yet the grid spends its 16 tries on only four different learning rates. Random search tries 16 different ones, so it lands closer to the sweet spot, near 0.04, every time. A Latin hypercube goes one step further and guarantees the 16 values of each setting are evenly spread, which made its worst run (8.496) better than random search's worst (8.526).

The gain is real but modest, a little over one per cent, and it carries to the held-back areas: the grid's winner scores 8.874 there, and the overall winner from all our searches scores 8.567. Tuning polishes a model; it does not rescue a poor one. The bigger lesson is about where to spend effort: when one setting matters far more than the others, a grid wastes most of its budget.

A Roundhay teenager who knows where a search budget is wasted can direct an AI's experiments instead of just watching them, and that starts with writing the code. The longer argument is in why teenagers should still learn to code in 2026.