Getting started with Equarith 1.0.3

Equarith is an offline symbolic-regression workbench. It searches for mathematical expressions that relate one numerical target column to one or more numerical input columns, then lets you inspect, validate, predict with, and export the resulting formulas.

Equarith performs data preparation, search, analysis, prediction, project storage, and formula export on your computer. It is designed for numerical regression: it does not train classifiers, and a discovered relationship is not evidence of causality.

Workflow at a glance

Import a delimited numerical file, paste a table, or open a generated sample.
Check column names, units, missing values, malformed values, and obvious data-quality problems.
In Search, select one target and the input variables that may explain it.
Choose the objective, train/test split, allowed functions, structural limits, and stop budget.
Start the search and watch the Pareto front improve.
Compare promising formulas on training and held-out test rows, then inspect their charts and residuals.
Copy a formula in the syntax you need, export a complete source file, or use it for custom predictions.
Save an .equarith project when you want to preserve the session, solutions, and a compatible checkpoint.

Quick-start tutorial: from a dataset to an exported formula

This tutorial uses a regular CSV, TSV, or delimited numerical text file. The first row should normally contain column names, and the table must contain at least two numerical columns and three rows. One column will be the value to predict; the others are candidate inputs.

1. Load the dataset

Equarith normally opens with the generated Braking Distance dataset already loaded. To use your own file, select Import File in the Input workspace, use File > Import Dataset…, or press Ctrl+I on Windows/Linux or Command+I on macOS. If the welcome screen is visible, Open Dataset starts the same import workflow.

Choose the data file, then review the bounded import preview. Equarith detects the delimiter, header, decimal style, character encoding, and quote character. If the preview is wrong, correct those controls before continuing. Check that:

the preview has the expected number of columns;
the first data row has not been mistaken for a header, or vice versa;
decimal commas or decimal points are interpreted correctly; and
the default missing-value tokens—an empty field, NA, and NaN—have the intended meaning; another token such as ? is reported as invalid.

Select Import. Equarith validates the complete file before replacing the active dataset. The Input workspace then shows the editable grid and data-quality summaries. Importing never changes the source file.

2. Check the data

Confirm that the target and candidate input columns are numerical and have meaningful, distinct names. Formula symbols are derived from these names. Correct obvious typing errors, and inspect the missing and invalid counts before searching.

Rows with a non-finite value in the selected target or any selected input cannot participate in that search. Numerical identifiers, dates, or category codes should only be used as inputs when their numerical order and distance are meaningful. Keep units consistent within each column.

3. Configure a first search

Open the Search workspace. The highlighted configuration sections show anything that still needs attention.

In Variables:

Select the output column under Target variable.
Select one or more explanatory columns as inputs. The target is automatically excluded from the normal input list.

For a first run, use these starting points:

Objective metric — RMSE. This is a general-purpose error in the target's unit; lower is better.
Train / test split — 80 / 20, Random. Search uses 80 percent and reports held-out behavior on 20 percent. Keep the split seed to repeat the partition.
Functions — select Defaults explicitly. This applies the recommended conservative mixture of arithmetic and protected elementary functions. Add functions only when the domain justifies them.
Maximum formula complexity — default, 65. This prevents formulas from growing without bound. Lower it when interpretability is more important.
Optimize coefficients (LM) — enabled; Optimization preset — Balanced. This fits numerical constants without spending all the budget polishing each structure.
Evaluation strategy — Auto. Equarith uses full evaluation or progressive filtering according to the dataset. Published scores still use all training rows.
Normalization — leave it at its current default for the first run. It can improve numerical conditioning; published formulas are converted back to source units.
Time limit (optional, minutes) — enable it, then enter a short limit for a trial. It is disabled by default. Search stops cleanly when the enabled budget expires. A longer run explores more candidates but does not guarantee a better model.

The function set controls what forms Equarith is allowed to discover. The metric controls what “better” means on training data. Complexity is a separate Pareto objective, so Equarith does not collapse accuracy and simplicity into one hidden score.

If Start remains unavailable, read the validation message beside the highlighted section. Common causes are a missing target or input, too few eligible rows, an impossible split, no selected function, or incompatible structural limits.

4. Start and monitor the search

Select Start in the command bar, use Search > Start Search, or press F5. Search runs in the background. The status area reports elapsed time, evaluated candidates, evaluation rate, and the current best objective.

Improving formulas appear in the Pareto front table. Each row balances training objective and expression complexity; a test objective is also available when the split includes held-out rows. The row with the smallest training error is not automatically the best practical formula. Look for a useful elbow where a large increase in complexity buys only a small improvement.

Select Stop or press Shift+F5 when you have enough results. The stop is cooperative, keeps published solutions, and writes a checkpoint when possible. A configured time limit uses the same orderly stop path.

There is no separate Resume button. When a compatible checkpoint exists after a stop or project reload, Start resumes it automatically. Changing checkpoint-sensitive search configuration invalidates that continuation and Start begins a new search instead.

5. Select a result and use the charts

Select a row in the Pareto table. The detail area offers Fit, Observed vs Predicted, Residuals, Pareto, and Search History.

Start with Fit to compare the observed target with the selected formula. Then check Observed vs Predicted for distance from the ideal one-to-one relationship and Residuals for curves, bands, widening spread, or isolated failures.

The chart interaction is the same throughout Equarith:

point at a plotted value to read its details;
scroll or pinch to zoom around the pointer;
drag with the primary mouse button to pan;
hold Shift and drag a rectangle to zoom into an area;
double-click the plot, select Fit or Reset, or press Home or 0 to show all visible data again;
use + and - to zoom and the arrow keys to pan when the chart has keyboard focus; and
use Grid, Lines, Points, scale, axis, and legend controls to change the view without changing the formula or its scores.

The chart toolbar can copy the image to the clipboard or export a PNG and the plotted series as CSV. Large charts may draw a bounded representative set of points for responsiveness; this does not reduce the rows used to calculate the published metrics.

6. Copy or export the formula

To copy one result, right-click its row in the Pareto front:

Copy formula copies the ordinary Equarith expression.
Copy as opens a submenu for LaTeX and 24 scientific, software, and industrial syntaxes, including Python, R, Julia, MATLAB, C, C++, Java, C#, JavaScript, TypeScript, Lua, Excel, PostgreSQL, and IEC 61131-3 Structured Text.
Export formula… writes the selected solution to a file. The Export Solutions… command can write the complete Pareto front.

Copy as puts only the selected expression in the clipboard. It is ideal for a notebook, document, source expression, or quick test. A protected operation may require helper definitions in executable code; use Export formula… when you need the complete standalone module, helpers, stable parameter mapping, and metadata. CSV, JSON, and plain-text solution reports are also file exports rather than Copy as targets.

Always test generated code against Equarith on representative, boundary, and deliberately invalid inputs before deployment.

7. Save the work

Use File > Save As… to create an .equarith project. A project can preserve the dataset or a verified reference to it, search settings, published solutions, and a compatible search checkpoint. Save again after choosing the final formula or changing important settings.

Try a generated sample instead

If you do not yet have a file, keep the Braking Distance dataset loaded at normal startup, or choose Load Example in the Input workspace. The welcome screen also offers it under Try a sample. Other generated samples are Free Fall, Electricity Cost, Projectile, and Logistic Growth. They follow the same Search, chart, and export workflow described above.

Demo license limits

An unactivated installation runs in Demo mode. The full application remains usable, but each search uses at most:

the first 200 source rows; and
four selected input variables.

Additional rows and columns remain visible, editable, importable, exportable, and savable. The Search workspace identifies which data will not participate in a Demo search. Academic and Pro licenses remove these two licensing limits; normal memory and engine safety limits still apply. The headless CLI uses the same license state and limits under the same system account.

Before relying on a formula

Keep a test set that did not guide selection whenever the dataset is large enough.
Inspect invalid predictions and domain-sensitive operations such as strict division, logarithm, and square root.
Check units and plausible behavior outside the observed range.
Treat correlation and good fit as descriptive evidence, not proof of causality.
Record the dataset, version, objective, split seed, search seed, functions and costs, worker count, and stop budget when reproducibility matters.
Validate the final formula on data from the real deployment regime.

Complete documentation

Data and setup:

Search:

Results and reuse:

Projects and preferences:

Automation, privacy, and support:

Equarith 文件。