Spread of the prediction (standard deviation) [MPa]
- Tap the plot to plan five tests.
- Make the largest spread as small as you can.
- Let the model plan: where does it test first?
Planned tests
Largest spread now
With the settings fixed, the spread depends only on where you test, not on the results.
- Tap the plot to drill. Find the richest spot.
- After five boreholes, show the true grade.
- Let the computer drill: compare upper confidence bound and greedy.
Boreholes left
Best so far
The computer's rule
- Make the length scale very short: what happens between the tests?
- More test scatter: why does the band stay wide at tests?
- Switch on Abrams' law: what changes outside the tests?
Spread at the test 1.30
Spread between the tests, at 0.93
More options
Tables and notes for all three steps.
Table: the model's plan, test by test
| Test | Water / cement | Largest spread before it [MPa] |
|---|
Table: the four tested mixes
| Cement [kg/m³] | Water [kg/m³] | Water / cement | Strength after 28 days [MPa] |
|---|
Table: the boreholes
| Borehole | Position [m] | Grade [g/t] | Drilled by |
|---|
Table: every rule with five boreholes
| Rule | New boreholes [m], in order | Best [g/t] |
|---|
Concrete data: I-C. Yeh (1998), Cement and Concrete Research 28, 1797–1808; 1030 concrete mixes from several laboratories (UCI Machine Learning Repository). The four tested mixes are real mixes of this data set, tested after 28 days; Abrams' law is fitted to all 425 mixes tested after 28 days. The gold line is made up for this demo, as in the lecture notes.