Raw results and code of a benchmark of six NEAT variants on 20 tabular medical datasets

Six algorithms were compared under one protocol: NEAT, FS-NEAT, NEAT-DBM, HyperNEAT, C-NEAT and L-NEAT, all implemented in the PolyNEAT library. Each algorithm was run 15 times on each of 20 public medical datasets, with 150 generations, a population of 50 and a stratified 66/34 train/test split. Run i evolves with seed 42 + i and splits the data with seed 1042 + i, hence all six algorithms of a run use the same split. The Parkinson's dataset is split by subject; all other datasets are split by row. The record has two archives: code_and_data.zip - archive with code that was used for gathering data: poly-neat directory consists of: polyneat - library with the configurations of all variants, gpu_sweep - the experiment runner, paper_analysis - the scripts that compute the tables and figures of the paper, reproduction instructions.md - how to rerun the experiment, data directory consists of: runs.csv - one row per run with metrics, seeds and runtime fitness_history.csv.gz - best training fitness in each generation sweep_meta.json - settings, software versions and SHA-256 checksums of the dataset files results_full.zip - the complete raw output of all 1800 runs: per-run records with the best fitness and number of species in every generation, network outputs and predictions on the training and test sets, the train/test indices of every split, the best evolved network of every run, and SHA-256 checksums of all files. The supplementary Tables S1-S4 of the paper are included as PDF files. The datasets are downloaded automatically through public sources: UCI, KEEL, scikit-feature and, for Pima, the jbrownlee/Datasets mirror on GitHub.