Symbolic regression engine
Closyr breeds thousands of candidate expressions, keeps the ones that fit, and hands you a formula you can read — not a black box. It runs in your browser on WebAssembly, or on a server when the data gets big.
The output is an expression — 0.5x² + sin(x) — not a fitted
black box. Every candidate comes with its complexity, so you can trade
accuracy for something you can actually publish.
The whole engine compiles to WebAssembly. Turn on In browser and your data never leaves the tab — no upload, no account, no server.
Dozens of populations evolve independently and trade their best members, which keeps diversity up and stops the search settling into one local optimum.
Expressions are compiled to flat bytecode and scored in SIMD batches, so the search spends its time exploring instead of interpreting trees.
Every promising structure is handed to a Levenberg–Marquardt pass that fits its numeric constants, so a good shape isn't thrown away for a bad coefficient.
The equation re-typesets as the search improves it, next to a live fit plot and the accuracy-versus-complexity front. Pause, resume, or stop whenever you've seen enough.
Pick a benchmark dataset, upload a CSV, or draw the curve you're thinking of with the brush tool. One to fifteen input variables.
Each population starts from a different family of expressions — polynomial, trigonometric, exponential — and mutates, breeds, and prunes from there.
Good expressions migrate between populations. A hall of fame keeps the best expression found at every complexity, not just the best overall.
The run ends with a set of solutions trading accuracy against size. Click any point on the front to see that expression plotted against your data.