Isotonic Regression: Pool Adjacent Violators, Monotone Least-Squares Fit and Block Merging
Fit the best non-decreasing curve to noisy data by pooling adjacent violators into blocks. Change the noise, sample size and the true shape, watch blocks merge from left to right, and compare the isotonic fit with a straight line and the raw data. A Blender-built plotting easel, with numerical checks of the min-max closed form, mass preservation, optimality and the KKT conditions built in.
Every keyword search engine has to answer the same question millions of times a second: given a query, which of billions of documents deserve to be shown first? BM25 is the classical answer — a probabilistic ranking function that balances how often a query term appears in a document against how rare that term is across the whole corpus, while penalizing documents that only look relevant because they are unusually long. This simulator renders a synthetic 18-document corpus in 3D, one bar per document, and recomputes every document's BM25 score live as you pick query terms and drag the k1 (term-frequency saturation) and b (length-normalization) constants — the same two knobs that Elasticsearch, Lucene and Solr expose to real search engineers.
Fit the best non-decreasing curve to noisy data by pooling adjacent violators into blocks. Change the noise, sample size and the true shape, watch blocks merge from left to right, and compare the isotonic fit with a straight line and the raw data. A Blender-built plotting easel, with numerical checks of the min-max closed form, mass preservation, optimality and the KKT conditions built in.
3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install