v4.0
November 8, 2026
This version brings a ~3x faster search engine and an improved search algorithm, which together find much more accurate formulas. There is also a new UI, a scikit-learn estimator for the Python library, and over 30 bug fixes.
Changes
- ~3x faster search engine: The search now goes through ~3x as many formulas per second per CPU thread as v3.3.1 on Linux and macOS, and 4.6x as many on Windows (up to 8x depending on the dataset).
- Much more accurate formulas: The faster engine and an improved search algorithm together find much more accurate formulas in the same search time. On the SRBench benchmark, with 1-minute searches, 4.0 left 2.6 times less unexplained variance than v3.3.1 on datasets with no known formula, and found the exact formula almost twice as often on physics problems with a known answer.
- New UI: The whole UI has been redesigned, with a cleaner look and a more compact layout. The top bar shows the dataset's name and size, the tabs have icons, a new status bar shows the progress of the search and the best error, and the selected solution has a "Copy formula" button that copies it with all its digits. The Input tab is now called Data.
- Large datasets in the UI: Datasets with tens of thousands of input variables no longer slow down the UI: 50,000 input variables now load in ~1 second, while v3.3.1 crashed. The spreadsheet is also a lot faster: on a 100,000-row dataset, editing a cell takes 37 ms instead of 766 ms.
- scikit-learn estimator: The Python library now includes
TuringBotRegressor, a scikit-learn compatible estimator:TuringBotRegressor(max_time=30).fit(X, y).predict(X_new). The formulas found, from simplest to most accurate, are kept inequations_, andsympy()returns the selected one as a sympy expression. - Weighted metrics: You can now use the columns of your dataset inside the aggregations of a custom search metric, for example a weight column
winsum(w*pow(actual-predicted,2))/sum(w). Aggregations can also be nested, for weighted versions of metrics like R² or correlation. - Linux AppImage: TuringBot is now also available as an AppImage, a single file that runs on most distributions without installation. The AppImage's command-line mode also works on servers without a graphical environment (on machines without FUSE, like Docker containers, run it with
--appimage-extract-and-run). - Universal macOS installer: A single .dmg now runs natively on both Apple Silicon and Intel Macs, so there are no more separate downloads to choose from.
- Make it possible to model a single time series from its own past: on a dataset with only the target column and "Target variable in history functions" enabled, the search only returned constants, and now finds formulas like
delay(3.9*(y-y*y),1)for the logistic map. - Show a warning when the input file contains values that aren't plain numbers, like
$30.00, dates or text, with how many there are and how the first one was read. Previously, they were silently read as 0 or as their leading digits. - Print formulas with fewer redundant parentheses and signs, like
-0.5*x+3instead of(-0.5)*x+3, and write their constants so that they read back as exactly the same number. - Load large input files about twice as fast.
- Reduce the CPU usage of the UI during searches by ~40%.
- Make the Prediction tab plot the selected solution right away, with one of its variables set to Range over its full observed range, with up to 50 steps instead of 5.
- Make Ctrl+Arrow, Ctrl+Shift+Arrow and Ctrl+A in the spreadsheet work like Excel.
- Remember the size, position and maximized state of the window between launches, and keep the window inside the visible area of the screen.
- Only print the formulas in command-line mode when they change, instead of every second.
- Python library: Make searches start sooner, and work on a Mac accessed over SSH.
- Windows only: Make the installer smaller, and show a message instead of installing on Windows versions older than 10, which aren't supported.
Bug fixes
- Fix a bug where the command-line mode, and therefore the Python library, could freeze on Windows and macOS when running with more than 1 thread (the default), so upgrading is highly recommended!
- Fix TuringBot failing to launch on Apple Silicon Macs running macOS 11, 12 or 13, and on older Linux distributions like RHEL/Rocky/Alma Linux 8, Debian 10 and Ubuntu 18.04.
- Fix numbers written as
+3,007or.5being read as 0. - Fix the spreadsheet losing data: "Apply changes" and "Export CSV" could drop rows or columns when the data had empty cells, rows deleted by clearing their cells were kept as rows of zeros, and "Export CSV" only exported the first 100,000 rows and 1,000 columns. "Apply changes" now also asks for confirmation before replacing a larger dataset with the part shown in the spreadsheet.
- Fix errors and plots being shown in normalized units when "Normalize the dataset" is enabled, and fix the classification, relative error, RMSLE and custom metrics with that option.
- Fix the program crashing, hanging or loading with shifted column names when a file had no complete row, or a header that didn't match the number of columns. Files that aren't plain text, like .xlsx spreadsheets, are now refused with an error message, and files with non-UTF-8 column names, like
temp°C, now load. - Fix a security issue where a crafted column name or custom formula, for example in a data or settings file someone sent you, or a Python library parameter, could run code on the computer.
- Fix
x*-ybeing read as a different formula in custom formulas, custom metrics and formula files. Formulas that can't be read exactly, likepow(a), are now refused instead of being silently changed. - Fix "Save settings" and "Load settings" changing the settings of a running search, which could crash it.
- Fix the program crashing when starting a search after a custom formula failed to parse, and when a custom input contained
f(), as iny = f(x, cos(z - f())). Also fix searches being refused with "History size error" because of the custom formula of a previous search. - Windows only: Fix the program crashing when another program, like Excel, had the output or predictions file open, and searches continuing to run in the background after the app or Python script that started them was closed.
- Windows only: Fix files with emoji or rare characters in their names failing to load or save.
- Fix custom formulas not accepting functions that are not selected under "Functions", like
roundiny = round(f(x)), a regression introduced in v3.1. - Fix formulas loaded with "Load formulas from file" or
--formulas-filesometimes not matching the saved formulas, especially with "Normalize the dataset" or custom inputs. - Fix formulas loaded with "Load formulas from file" or
--formulas-fileignoring the search constraints, like "Force solutions to include all variables" and the limits on the number of constants. - Fix some formulas being shown with missing parentheses in some searches, like those with custom inputs.
- Fix the custom search metric giving wrong errors when it used a variable like
nin an aggregation and the custom formula had custom inputs, likey = f(x, x*z). - Fix settings that didn't do what they said:
--threads 0silently ran no search, a custom search with an empty formula ran a regular search, the "Residual error (%)" plot showed a fraction instead of a percentage, and settings files without anallowed_functionsline were refused by the command-line mode. - Fix "Force solutions to include all variables" and the "Number of distinct variables" limits miscounting variables, which could show solutions without some of the variables, or reject formulas with custom functions like
myf(x)+y. - Fix the main plot drawing the model as a zig-zag line when the input file was not sorted by the plotted variable.
- Fix the Python and C exports of formulas with
delay,moving_averageordecay, which called functions that were not defined, and the C export using the integerabs()function and not defining thegammaconstant. - Fix the number of formulas and the formulas per second being undercounted when using many threads.
- Fix file dialogs not opening in the last used folder on some Linux distributions, like Ubuntu 22.04 and Linux Mint 21.
- Python library: Fix
start_process(config=...)ignoring the search options, aninput_filegiven as apathlib.Pathbeing refused, temporary files being left behind after a failed start,refresh_functions()silently finding nothing when TuringBot exited with an error, and an empty console window opening on Windows when running scripts without a console, like.pywscripts. - Fix details of the formulas shown in LaTeX, like
1.5×10^{-07}instead of1.5×10^{-7}, and underscores in the names of custom functions making subscripts.
v3.3
February 24, 2026Changes
- Custom base functions: You can now define your own unary and binary base functions that participate in the symbolic regression search. Append
|costto set a custom complexity weight. Examples:bump(x)=exp(-x*x),mydist(x,y)=sqrt(x*x+y*y),gauss(x)=exp(-x*x)|2. Custom functions can reference previously defined custom functions. - Custom search metric: You can now define your own error metric formula using aggregation functions
mean(),sum(),median(),maxval(),minval()with the per-row variablesactualandpredicted. You can also combine built-in metrics likerms,mae,correlation, etc. in a single formula. Examples:mean(pow(actual-predicted,2))for MSE,sqrt(mean(pow(actual-predicted,2)))for RMSE,0.5*rms+0.5*maefor a weighted combination. - Custom constants: You can now define custom constants that participate in the search with a default complexity of 1. Append
|costto set a custom complexity weight. Example:g=9.81,c=299792458. - Custom inputs: You can now pass expressions as extra inputs in custom formula searches. Write
y = f(x, cosh(x*y)+2)and the search will use bothxandcosh(x*y)+2as input variables. Expressions can use any base function or constant, including custom ones. - Customizable function sizes: Set the size cost of each base function individually through the UI (numbers next to each function) or the CLI (
--function-size "sin:3 cos:3 pow:5"). - Performance optimizations based on profiling, increasing formula throughput by ~30% on small datasets.
- 3x faster file loading. A 768 MB file now loads in ~6.4 seconds, down from ~20 seconds, outperforming numpy's
loadtxt. - Optimize constants sampling in integer-only mode, as well as other optimizations to the core symbolic regression algorithm.
- Support resuming a previous search with "Load formulas from file" in custom search mode.
- Add
decay(f, r)as a new history function for exponentially weighted sums:decay(f, r) = f(x) + r*f(x-1) + r²*f(x-2) + ...where0 < r < 1. Complements the existingdelayandmoving_averagefunctions. - Python library: TuringBot is now available as a PyPI package. Install with
pip install turingbotand use directly withimport turingbot, no more manualsys.pathmanipulation or hardcoded executable paths. - Python library: The
input_fileparameter instart_process()now accepts numpy arrays and pandas DataFrames in addition to file paths. - Python library: The
pathparameter instart_process()is now optional. The library automatically detects the TuringBot executable at default installation paths for each OS and falls back to a PATH lookup. - Python library: The
allowed_functionsparameter now accepts a list of strings (e.g.["+", "sin", "cos"]) in addition to a single space-separated string. - Formulas, predictions, and exports now use the original data scale when "Normalize the dataset" is enabled, by wrapping outputs with
output * std + avg. - Make it possible to round-trip formulas with
--formulas-filewhen "Normalize the dataset" is enabled, by including raw formula columns in the output file. - When exporting solutions as Python or C, all constants (built-in like
pi,e,sqrt2, and custom) are now properly defined in the generated code. - Add an "Update available" notification in the UI when a new version of TuringBot is released, with a link to download it.
- CLI flags now take precedence over settings file values when both are provided.
- Allow relative paths for input files and all file flags.
turingbot ./input-file.csvnow works. - Suppress GPU and DBus warning messages in headless CLI mode, e.g. over SSH.
Bug fixes
- Fix the F-score (F-beta) calculation, which was using the user-provided beta directly instead of squaring it first.
- Fix the bound search mode labels being swapped: selecting the documented lower bound index in the CLI or config file actually performed an upper bound search, and vice versa.
- Fix the Prediction tab crashing when the selected formula contains history functions (
delayormoving_average). - Fix
delayandmoving_averagereturning NaN in the predictions output file. - Fix the Python library crashing when reading results from a partially-written output file, deleting user-specified output files on process termination, and not quoting file paths with spaces in command arguments.
v3.3.1
March 27, 2026- Improve search performance for scale-independent metrics (maximum relative error, Nash efficiency, log-loss, MCC, RMSLE).
- Improve search performance for custom formulas with complex left-side expressions like y*x + 2*sin(x) = f(x).
- Fix log-scaled plots being distorted due to the log transformation being applied twice.
v3.2
January 22, 2026Changes
- Add an "Export to CSV" button to export spreadsheet data.
- Optimize the spreadsheet by making the number of rows/columns grow dynamically.
- Align the binary cross-entropy (log-loss) calculation with scikit-learn by adding epsilon clipping. This prevents numerical errors when predicted probabilities are exactly 0 or 1.
- Remove unused fonts and assets, reducing startup time and installation size.
- Add
--activate,--deactivate,--license-statusflags for activating the program on headless environments.
Bug fixes
- Fix the Matthews Correlation Coefficient (MCC) calculation, which had (TN*FN) instead of (TN+FN) in the denominator -- sqrt((TP + FP) * (TP + FN) * (TN + FP) * (TN + FN)).
- Fix the menu icon flickering when hovered.
v3.1
May 26, 2025Changes
- Make a major optimization to the symbolic regression algorithm that helps it escape local minima more effectively.
- Add a new percentile error search metric with an option to set the quantile. This metric is useful for datasets with outliers.
- Add new options for constraining formulas: minimum and maximum number of constants, minimum and maximum number of different variables, maximum number of occurrences per variable.
- Add the option to include constants like pi, e, sqrt2, ln2 in the search.
- Make solutions with 'Normalize the dataset' directly usable by replacing variables with (variable-avg)/std in the output formulas.
- Add a plot to the Log tab showing record errors over time to visualize search progress.
- Add "New record!" log messages when solutions with new record-low errors are found.
- Improve logarithmic scales for plots, which now allow visualizing actual numbers (rather than transformed numbers) while only changing the axis scale itself. The symlog options had to be removed to support this change.
- Add a 'Plot type' dropdown to switch between regular plots and observed vs predicted plots.
- Make it possible to select and load multiple seed formula files at once.
- Rename 'cross-validation' to 'test-sample' throughout the UI, since this more accurately represents what is implemented.
- Add more diagnostic metrics to the Search tab: Mean error, Median error, Maximum error, Median relative error, Maximum relative error.
- Organize the search metrics dropdown with regression metrics at the top and classification metrics at the bottom.
- Reduce the number of log messages by not logging solutions with negligible error improvements.
- Add a confirmation message after finishing processing seed formulas from a file.
- Improve the formatting of error numbers in the Solutions table.
- Only show test sample settings (random seed, sequential vs random) when a test sample is selected.
- Only show the F-score beta parameter input when the F-score metric is selected.
Bug fixes
- Fix the UI sometimes crashing or becoming all white and unresponsive during searches.
- Fix the program failing to launch when two instances are launched simultaneously.
- Prevent seed formulas with disallowed functions from being loaded.
- Fix loading formulas containing spaces from files.
- Fix the R-squared calculation to match the most commonly used definition exactly.
v3.1.1
May 27, 2025- Fix "Maximum occurrences per variable" not counting the x in functions like sin(x).
- Show an error message instead of crashing when trying to load a badly formatted input file.
v3.1.2
June 1, 2025- Fix a crash during searches introduced in v3.1. This fix also improves search speed, so upgrading is highly recommended!
v3.1.3
June 18, 2025- Fix a rare crash during searches.
v3.1.4
July 3, 2025- Fix a crash when the number of generated formulas exceeds around 4e9.
- Fix the displayed elapsed time doubling when the search is paused.
- Fix custom searches with constants, like y = f(x) + pi.
v3.0
March 11, 2025Changes
- New UI layout with a left sidebar for navigation across tabs.
- New Input tab to enter data directly in the UI through a built-in spreadsheet.
- New Prediction tab to predict model outputs for fixed inputs or ranges of inputs, with a plot to make projections. The default values come from the last row in the dataset.
- Consolidate all search options in the Search tab and put them in collapsible sections.
- Make all plots interactive, with options to zoom in and out, pan, and save the plot. Also improve the plot styles.
- Improve the style for the selected solution in the solutions box to make it more readable.
- Add a "Show combined data (training + testing)" option for the main plot.
- Make the input field for the custom formula larger.
- Make the statistical summary area a table instead of raw text.
- Check all but the last checkbox under "Input variables" by default when loading a dataset.
v3.0.1
March 28, 2025- More readable log messages, with horizontal scrolling for long entries
- Make the "Normalize the dataset" option normalize the row number as well
- Fix spreadsheet compatibility with decimal numbers using commas (common when pasting from Excel)
- Fix an issue where spreadsheet would repeat pasted values when selection had more rows than content
- Fix UI crashes that caused gray screen in some cases after the 3.0 update
- Fix "Solution info" box not clearing when loading a new dataset
- Fix variable names containing "$" (like "$x") not rendering properly in the Solutions box
v3.0.2
April 11, 2025- Make it possible to resize the spreadsheet columns.
- Fix pasting data to the spreadsheet when no selection is made.
v2.23
February 6, 2025Changes
- Repeated terms in custom search: You can now use indexed function terms (e.g.,
f1(x),f2()) to find formulas with multiple occurrences of the same unknown function. Please consult the updated documentation for details. - UI: Increase the vertical space between solutions in the Solutions box for better readability.
v2.23.1
February 17, 2025- Make the program launch slightly faster.
- Fix the program becoming unresponsive after trying to start a search with an invalid custom formula.
v2.22
November 30, 2024Changes
- New command-line flags: New flags have been added to customize searches without a config file, including
--search-metric,--train-test-split, and--custom-formula. All UI settings can now be configured through these command-line flags. Runturingbot --helpor see the documentation for more details. - Python library updates: Added support for arguments like
search_metricandtrain_test_splitin thestart_processmethod, allowing programmatic control of all UI settings without a config file. See the documentation for details. Also added exception handling to raise errors for invalid paths. - History functions generalization: The history functions (
delayandmoving_average) have been updated to accept general functions in the first argument, enabling formulas likedelay(x*sin(x),2). The history functions are now also supported in custom formulas, such asy = delay(f(x),1)ory = delay(f(x),f()). - Improve the UI responsiveness during optimizations when using large input files, such as those containing millions of rows.
Bug fixes
- Fix a command-line mode issue where only the first column was used as input by default instead of all columns except the last.
- Fix cases where zombie processes were still left behind after the program was closed.
- Fix a bug that caused the program to hang when base functions appeared in custom formulas without being selected under "Search options".
- Fix an issue that prevented the program from working in command-line mode over SSH.
v2.22.2
December 30, 2024- Increase the maximum formula size to 1024 (from 500), and enforce this limit in the UI.
- Add
enterandexitmethods to the Python library for use in a "with" statement. - Add a
--versioncommand-line flag. - Minor aesthetic improvements to the UI.
- Fix the Python library leaving zombie processes when the script is interrupted by Ctrl+C, killed by closing the CMD window on Windows, or terminated with
kill -9on Linux.
v2.21
September 9, 2024Changes
- Add the option to specify a custom number of rows for the train/test split.
- Add the RMLSE (Root Mean Squared Logarithmic Error) metric. Similar to "Mean Relative Error," it handles target variables across multiple magnitudes but penalizes large errors less aggressively. It requires the target variable to be strictly positive.
- Add a y=x line to the "Observed vs Predicted" plot to better visualize the points.
- Add the option to choose between log and symlog scales for plots. Previously, plots were always scaled using symlog, which behaves oddly when values are between 0 and 1.
- When using cross-validation, export the training rows first, followed by the testing rows, in order, when clicking "Export data + predictions".
- Set the "niceness" level of the program to 19 by default on Linux/MacOS. This helps prevent the UI and system from becoming unresponsive when using all CPU threads on these OS.
- Toggle checkboxes when clicking on function names under "Search options".
Bug fixes
- Fix the program sometimes crashing when starting a new run.
- Fix incorrect output for "Export data+predictions" when the solution contains delay or moving_average.
- Fix moving_average persisting after starting a search with both delay and moving_average, stopping, and then starting a search with only delay.
- Prevent "Load formulas from file" from loading formulas containing base functions not allowed in the current search.
- Fix custom formulas not being parsed when a dataset variable has a name similar to a base formula, such as "y = tan(x) + f()" not parsing if one of the variables is named "ta".
- Fix the LaTeX output of logical_and and logical_or.
v2.21.1
September 19, 2024- Fix the program sometimes crashing when starting a new run.
- Python library: fix some macOS crashes, fix zombie processes being left behind if the script gets interrupted.
v2.20
August 15, 2024Changes
- Add a progress message to the UI when a very large file is being loaded.
- Do not truncate variable names to 10 characters. Instead, use the full variable names in formulas.
- Add the RSS (residual sum of squares) error metric.
Bug fixes
- Fix history functions (delay and moving_average) sometimes being shown with the wrong variable in the argument when the search contains more than one input variable.
- Fix a bug that caused the program take a lot longer than necessary to load large files (> 1 GB) in recent releases.
- Add missing "Correlation coefficient" option in the "Analysis" tab.
- Fix LaTeX rendering for the "moving_average" function.
- Fix the "Statistical summary" field in the "Analysis" tab overflowing when many input columns are present.
v2.19
July 19, 2024Changes
- LaTeX rendering: Equations in the "Solutions" box are now rendered as LaTeX, making them a lot easier to read and interpret.
- Export as LaTeX: This new menu option makes it possible to export equations in a format that can be directly copied and pasted into academic publications and technical reports.
- Less parentheses: A new algorithm for placing parentheses in equations has been developed, reducing their frequency and thus the character count of equations.
- Select/deselect groups of base functions by clicking on the corresponding headers. For instance, "Trigonometric functions" and "Exponential functions" are now clickable.
- Do not select "Row number" when clicking on the "All/none" button for input variables, as the row number is usually not a relevant variable.
- Minor UI optimizations to reduce its CPU usage.
Bug fixes
- macOS signing: TuringBot is now a signed macOS application, resolving the cryptic "TuringBot is damaged and can't be opened. You should move it to the Trash" error message.
- Add back the "Correlation coefficient" metric, accidentally removed in v2.18.
- Fix periodic solutions being saved with "_tmp" extension on Windows.
v2.18
June 10, 2024Changes
- Cross-platform experience: We have rewritten the UI from scratch in a new framework, ensuring a consistent and identical experience across all major operating systems.
- MacOS support: We have added new .dmg installers for macOS, available in both Apple Silicon (M1/M2/M3) and Intel versions. These updated installers do not require installing any dependencies manually and should "just work".
- Caching of plots and UI errors: Solution errors and plots are now only calculated once per solution, allowing you to go back and forth between different solutions on the list without wasting CPU cycles.
- Improved logging: We have streamlined "New solution found" log messages, grouping them together every 1 second for enhanced readability.
v2.18.1
June 16, 2024- Smoother startup: The program's launch has been refined to be visually smoother and less glitchy.
- Fix cross-validation errors not being correctly shown in the UI.
- Fix the program crashing when "Matthews correlation" was selected as the error metric.
- Windows only: Fix loading input files with UTF-8 characters in the filename, and fix saving settings when the username has UTF-8 characters.
- Fix loading "Integer constants only" setting from a settings.cfg file (it was being ignored).
v2.16
February 3, 2023Changes
- Implement dataset normalization option (Advanced > Normalize the dataset). This speeds up the search when the input values are large.
- Optimize the program for very large numbers of rows/columns. The interface should be much more responsive now and files with many columns should also load faster.
- When cross-validation is enabled, only show the corresponding points (train or test sample) in the plot.
- Reduce the minimum value of "Maximum formula size" from 10 to 1.
- Add the input file filename to the "Statistical summary" box.
- Make periodic output save immediately when the periodic output option is first set.
- Use >= instead of > for upper bound search, and <= instead of < for lower bound search.
- Stop the elapsed time counter from elapsing if the program is paused.
Bug fixes
- Make "recent files" work in the file selection dialogues.
- Fix bug where not all moving average lengths up to "Maximum history size" are used.
v2.16.1
February 21, 2023- Program still getting slow and unresponsive for large number of rows.
- Interface errors not updated after loading formulas from a file with the search paused.
v2.14
November 17, 2022Changes
- F-score: the F1-score metric has been replaced by the more general F-score metric with a customizable beta parameter. Lower values of beta favor precision over recall. Useful for stock market applications.
- log10(x) base function has been added.
- Random seed for random train/test split generation can now be manually set. When left at the default value (-1), a random seed will be used each time.
- ~30% faster performance after optimizations.
- Customizable interface size: the interface size can now be adjusted under menu > interface scale. By default, this scale is now set to 90%.
- Folder memory: the last used folders for exporting and loading files are now remembered even if the program is closed and launched again.
- Negative predictive value metric has been added to the "Solution info" box for completeness of the confusion matrix.
- Safe saving: whenever an output file is generated, it is now first saved with a _tmp extension and then renamed. This prevents the creation of partially written files in the event of power outages or crashes.
- Select/deselect all: history functions are no longer toggled on by this button and must now be toggled manually.
- macOS version: no longer requires manual dependency installations using brew. Feedback from macOS users as to whether TuringBot is working as expected would be highly appreciated.
- Input dialog: on Windows and macOS, the system dialog for selecting the input file has been replaced by a more practical internal dialog.
- Command-line improvements: it is now possible to load seed formulas from the command line with --formulas-file FILENAME and export formulas+predictions periodically with --predictions-file FILENAME.
- Python improvements: the same options above have been added to the start_process method: predictions_file=None, formulas_file=None.
Bug fixes
- Large formulas are not shown if more than 60 formulas with decreasing error are on display.
- Broken "output type" button under the "set periodic output" window on Windows 10.
v2.12
May 20, 2022Changes
- macOS version: a native macOS version of the program is now available. Feedback from Mac users would be highly appreciated.
- Analysis tab: this new tab contains a customizable "Error vs Size" plot as well as statistical information about the input data. It replaces the old Prediction tab, which most users did not consider useful.
- Binary cross-entropy: new search metric. Solves binary classification problems in terms of probabilities between 0 and 1. To use this metric, the target variable must contain two (and only two) classes represented by the numbers 0 and 1.
- Matthews Correlation Coefficient: new search metric. Regarded to be one of the best classification metrics, it combines true positives, true negatives, false positives and false negatives into a single number between -1 and 1. See the documentation for details.
- Precision, Recall, and Sensitivity: these metrics have been added to the "Solution info" box in the main tab of the interface. They allow solutions obtained with F1 score/MCC metrics to be much better interpreted and analyzed.
- min(x, y) and max(x, y): new base formulas.
- Export data+predictions: this option has been added to the menu, allowing data and predictions to be manually exported at any time.
- Interface improvements: menus have been made more user friendly (no more overflowing menus with white space) and the interface colors have been made slightly more readable.
- Periodic output: outputs are now only saved if new solutions have been found, preventing unnecessary disk usage.
- File reading: input files are now read 30% faster.
Bug fixes
- Broken interface on the latest Ubuntu versions.
- Optimization considers y = 1e-200 to be a perfect fit when the correlation coefficient metric is used.
v2.10
February 18, 2022Changes
- Observed vs predicted plot: you can now see your predictions as a function of the target variable by clicking on Plot x axis > Observed.
- Symlog plot: when selecting log y or log x in the "Plot scale" menu, the behavior is now to calculate the "symlog" of the data just like in Python's Matplotlib, allowing negative numbers to be visualized nicely.
- Folder memory: if you select a file in a menu, and then open that same menu later, the folder shown is now remembered.
- Update frequency: the interface now updates more frequently.
- Input file reading: now 30% faster, more robust, and with error handling: if the file is empty or unreadable, the program now shows an error message instead of crashing.
- Special characters: the program now supports reading/saving files with UTF-8 filenames (including accented characters and non-Latin alphabets like Chinese and Russian).
- Decimal places: when exporting solutions as text/Python/C or exporting data, constants are now written with more decimal places to avoid loss of accuracy.
- Error comments: when exporting solutions as Python/C, a comment is now shown above each solution with its corresponding error.
- Cross-validation error exporting: when saving solutions as text, the cross-validation error is now also saved (if cross-validation is enabled).
- Cross-validation optimization: optimizations with cross-validation enabled are now much faster, especially for large input files.
- Command-line help: running turingbot (or TuringBot.exe) with --help flag now shows a help message.
Bug fixes
- Fix bug that caused the program to randomly crash after running for several hours.
- Fix "Plot x axis" and "Plot scale" menus extending beyond the program window and looking weird.
- Fix ghost solutions appearing in the Log messages.
- Fix "Set periodic output" dialog disappearing after being closed once.
- Fix broken function declarations when exporting as C/Python in Advanced mode.
- log2(x) is now correctly read when typed into Advanced mode.
- R^2 is now 1 and not 0 for perfect fits.