Release History

Version 0.4.6 - 6 Apr 2026

  • GPU/CUDA Acceleration: Introduced high-performance GPU-accelerated Genetic Programming. This project is a CUDA-enabled fork of the original gplearn by Trevor Stephens.

  • Hybrid VM Interpreter: Implemented a static CUDA C++ Virtual Machine interpreter that eliminates JIT overhead during evolution, delivering significant speedups on massive datasets.

  • Batched Evaluation: Offloaded entire population evaluation to the GPU in a single grid launch.

  • Memory Optimization: Automated X-matrix transposition for coalesced memory access on NVIDIA hardware.

  • Library-wide Support: GPU acceleration is now available for SymbolicRegressor, SymbolicClassifier and SymbolicTransformer.

  • Packaging Update: The base package remains CPU-only, while optional extras now provide explicit pip install paths for cuda13 and cuda12.

  • Dependency Update: The cuda and cuda13 extras install cupy-cuda13x[ctk] and the cuda12 extra installs cupy-cuda12x[ctk] so common CUDA runtime libraries are pulled in automatically for pip-based installs.

  • CUDA Reliability: Fixed NVRTC compilation of the shared VM module by including the CUDA half-precision header needed by the mixed-precision kernel.

  • GPU Evolution Fix: Preserved feature-count metadata during batched GPU population generation so GPU-only program construction remains valid.

  • Transformer Fallback: SymbolicTransformer(device='cuda') now falls back to a NumPy correlation step when CuPy’s BLAS backends are unavailable, which is especially relevant for some Windows Python 3.14 environments.

  • Compatibility Documentation: Documented Python 3.11+ support and the currently verified Windows CPU/CUDA version matrix in the user docs.

Version 0.4.3 - 6 Jan 2026

  • Update dependencies to ensure compatibility with scikit-learn and newer versions of Python. scikit-learn 1.8.0 or newer will be required as well as Python 3.11 or newer. Also migrated packaging to hatch through this is unlikely to affect most users.

Version 0.4.2 - 3 May 2022

  • Require keyword only arguments for all public methods and functions to comply with scikit-learn SLEP009.

  • Replace n_features_ attribute with n_features_in_ to comply with scikit-learn SLEP010.

  • Update test suite to ensure compatibility with scikit-learn. scikit-learn 1.0.2 or newer will be required due to recent changes in their testing requirements. Also requiring joblib to 1.0.0 or newer to align with next release of scikit-learn.

  • Added the class_weight parameter to genetic.SymbolicClassifier allowing users to easily compensate for imbalanced datasets.

Version 0.4.1 - 1 Jun 2019

  • Fixed a bug with multi-processing and custom functions, allowing pickling of models with custom functions, fitness metrics or classifier transformers. joblib 0.13.0 or newer required in order to take advantage of this release in order to wrap functions for pickling saved models.

Version 0.4.0 - 23 Apr 2019

  • Added the genetic.SymbolicClassifier to use symbolic regression to solve binary classification problems. This passes the outputs of a program through a sigmoid function in order to translate the result into a probability of either class.

  • Allow users to express feature names as strings rather than X0, X1, etc. Graphviz and print() output can now be customized by setting feature_names=[...] in genetic.SymbolicRegressor or genetic.SymbolicTransformer.

  • Allow users to exclude constants from their programs by setting const_range=None in genetic.SymbolicRegressor or genetic.SymbolicTransformer.

  • Record details (similar to the verbose output) of the evolution in the estimator attribute run_details_ dict in genetic.SymbolicRegressor and genetic.SymbolicTransformer.

  • Pearson and Spearman correlation coefficients added as first-class metrics to genetic.SymbolicRegressor. These metrics allow for evolution of value-added features for second-stage estimators.

  • Added a low_memory parameter in genetic.SymbolicRegressor and genetic.SymbolicTransformer which can reduce memory use for cases where there are large populations or many generations by removing early generation program information. By Bartol Karuza and wulfihm.

  • Drop support for Python 2.7 and Python 3.4 to ensure compatibility with scikit-learn. scikit-learn 0.20.0 or newer will also be required due to recent changes in their testing suite. Additionally joblib 0.11 or newer will be required due to scikit-learn devendoring it.

Version 0.3.0 - 23 Nov 2017

  • Fixed two bugs in genetic.SymbolicTransformer where the final solution selection logic was incorrect and suboptimal. This fix will change the solutions from all previous versions of gplearn (and original gplearn). Thanks to iblasi for diagnosing the problem and helping craft the solution.

  • Fixed bug in genetic.SymbolicRegressor where a custom fitness measure was defined in fitness.make_fitness() with the parameter greater_is_better=True. This was ignored during final solution selection. This change will alter the results from previous releases where greater_is_better=True was set in a custom fitness measure. By sun ao.

  • Increase minimum required version of scikit-learn to 0.18.1. This allows streamlining the test suite and removal of many utilities to reduce future technical debt. Please note that due to this change, previous versions may have different results due to a change in random sampling noted here.

  • Drop support for Python 2.6 and add support for Python 3.5 and 3.6 in order to support the latest release of scikit-learn 0.19 and avoid future test failures. By hugovk.

Version 0.2.0 - 30 Mar 2017

  • Allow more generations to be evolved on top of those already trained using a previous call to fit. The genetic.SymbolicRegressor and genetic.SymbolicTransformer classes now support the warm_start parameter which, when set to True, reuse the solution of the previous call to fit and add more generations to the evolution.

  • Allow users to define their own fitness measures. Supported by the fitness.make_fitness() factory function. Using this a user may define any metric by which to measure the fitness of a program to optimize any problem. This also required modifying the API slightly with the deprecation of the 'rmsle' error measure for the genetic.SymbolicRegressor.

  • Allow users to define their own functions for use in genetic programs. Supported by the functions.make_function() factory function. Using this a user may define any mathematical relationship with any number of arguments and grow totally customized programs. This also required modifying the API with the deprecation of the 'comparison', 'transformer' and 'trigonometric' arguments to the genetic.SymbolicRegressor and genetic.SymbolicTransformer classes in favor of the new function_set where any combination of preset and user-defined functions can be supplied. To restore previous behavior initialize the estimator with function_set=['add2', 'sub2', 'mul2', 'div2', 'sqrt1', 'log1', 'abs1', 'neg1', 'inv1', 'max2', 'min2'].

  • Reduce memory consumption for large datasets, large populations or many generations. Indices for in-sample/out-of-sample fitness calculations are now generated on demand rather than being stored in the program objects which reduces the size significantly for large datasets. Additionally “irrelevant” programs from earlier generations are removed if they did not contribute to the current population through genetic operations. This reduces the number of programs stored in the estimator which helps for large populations, high number of generations, as well as for runs with significant bloat.

Version 0.1.0 - 6 May 2015