AI and the World's Hard Problems · Theoretical computer science · arXiv · 2026-10-23
Direct Inference on Admissible Sets (Diona)
The authors propose Diona, a regression method that estimates a state, its derivatives, and their admissible bounds directly from samples at a query point, without long global training. On a benchmark of six functions in three to eight dimensions, it was more accurate than a Gaussian process on noiseless data, especially with denser data, and like the GP its bounds contained the true value at more than 95% of query points for four functions. Its time to a first prediction after a data update was three to four orders of magnitude shorter than the GP's and two to three orders shorter than a neural network's, an advantage that holds when relatively few predictions are needed between updates.