Puzzle Solving without Search or Human Knowledge: An Unnatural Language Approach

TL;DR

This paper demonstrates that GPT-2 can learn to solve complex puzzles like mazes, Rubik's Cube, and Sudoku solely from text archives, without search or human heuristics, by fine-tuning on solved game data.

Contribution

It introduces a novel approach of using transformer models trained on text-archived game solutions to solve puzzles without search or domain-specific heuristics.

Findings

Transformer models can learn puzzle-solving strategies from text archives.

The method achieves solutions in environments with sparse rewards.

It bypasses traditional search and heuristic methods for puzzle solving.

Abstract

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 ($>10^{19}$) for the games provides a puzzle environment in which the solution has few intermediate rewards and a final move that solves the challenge.