The application of Generative Pre-trained Transformer (GPT-2) to learn text-archived game notation provides a model environment for exploring sparse reward gameplay. The transformer architecture proves amenable to training on solved text archives describing mazes, Rubik's Cube, and Sudoku solvers. The method benefits from fine-tuning the transformer architecture to visualize plausible strategies derived outside any guidance from human heuristics or domain expertise. The large search space (>1019) for the games provides a puzzle environment in which the solution has few intermediate rewards and a final move that solves the challenge.
@article{arxiv.2109.02797,
title = {Puzzle Solving without Search or Human Knowledge: An Unnatural Language Approach},
author = {David Noever and Ryerson Burdick},
journal= {arXiv preprint arXiv:2109.02797},
year = {2021}
}