Puzzle Solving without Search or Human Knowledge: An Unnatural Language Approach
This addresses puzzle-solving for AI systems by eliminating the need for search or human knowledge, though it appears incremental as it adapts existing transformer models to new domains.
The paper tackled the problem of solving puzzles with sparse rewards by applying GPT-2 to learn from text archives of games like mazes, Rubik's Cube, and Sudoku, achieving a method that visualizes strategies without human guidance in large search spaces exceeding 10^19.
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.