Flowcheck: A Discrete Flow Matching Approach for Generating Chess Configurations

• Vinícius do Carmo Melício UFABC
• Manolo Canales Cuba UFABC
• Charles Morphy D. Santos UFABC
• João Paulo Gois UFABC

Resumo

We introduce FlowCheck, a novel method for generating realistic chess positions by treating the board as a discrete grid. Our approach uses Discrete Flow Matching (DFM) to train a Vision-Transformer-based model that generates board states conditioned on high-level strategic labels, such as material balance and engine score. A key innovation is the ability to create positions based on specific tactical concepts and a post-processing heuristic to refine the outputs for greater accuracy. FlowCheck produces configurations with a repetition rate of less than 0.02%, and our refinement step significantly improves the fidelity to the desired engine evaluation. This approach offers a powerful tool for procedural content generation, with applications in tailored player training, strategic exploration, and the creation of unique educational content.

Palavras-chave: Training, Graphics, Technological innovation, Procedural generation, Accuracy, Databases, Safety, Engines