This repository contains our approach of solving the challenge posted on Kaggle back in the days.
The whole algorithm is described on Medium in great detail. You can find accompanying code in ga_tutorial.ipynb.
Almost equivalent code is used in GeneticSolver, with multiprocessing version of fitness scoring added. MPGeneticSolver goes one step further by running multiple GeneticSolver's in parallel and returning the best scoring solution across all of the cores. Cythonized version of make_move is written in life.pyx, which is responsible for ~8x speedup in fitness scoring, and thus the overall performance.
Genetic Algorithm have yielded 0.0634 score on private test set, which beats top solutions by a wide margin.
Another approaches were tried, including Random Forest and Neural Networks on sliding windows, with top score ~0.12034.
- Andrew Stadnik (@anstadnik)
- Vlad Paladii (@samaelxxi)
- Pavel Tyshevskyi (@ptyshevs)