Earlier today the marine-hull programme reported that a small wing-sail drone beating to windward in real harbour wind loses 12 points of progress against a clairvoyant controller, and that no forecast, however good, recovers any of it. Both statements are wrong, and one new test showed why.
The clairvoyant dynamic programme let the boat sit on the edge of its survey corridor and keep full speed. Every real controller had to tack there and pay a 30 second penalty. With the same rule applied to the clairvoyant, its score falls from 99.7% to 92.7% of the bound in July, so about 7 of the 12 points were never available. The honest gap to the header-following controller is 4.5 to 5.8 points.
The second error was the planner. The forecast experiments used a finite horizon with zero terminal value. A full-day value table built on wind smoothed to 15 minutes, then run against the true wind, closes 93 to 95% of the honest gap, about 5 points over header-following at the same tack count. Smoothing to an hour keeps about 38% of the gain, and two hours keeps almost none.
What it does not show: the controller here knows the whole day's slow wind exactly. A real forecast has error that grows with lead time, which is the next step. Numbers, code and the corrected log are in research/marine-hull (milestone 5c).