Comparative Analysis of GAT and BERT for Human-Like Playtesting
Automated playtesting is essential for modern game development, allowing developers to balance difficulty and ensure player engagement without the high costs of manual testing. While traditional methods like Convolutional Neural Networks (CNNs) have been used to predict player moves, they often struggle to adapt to new game mechanics or complex board layouts. This paper explores more flexible, general-purpose architectures—specifically BERT and Graph Attention Networks (GAT)—to better capture the relational structure of puzzle games like Candy Crush Saga and improve the accuracy of automated playtesting.
Moving Beyond Grid-Based Limitations
Traditional CNNs process game boards as fixed grids, which makes it difficult for them to understand relationships between non-adjacent tiles, such as those connected by portals. Furthermore, CNNs often require significant manual feature engineering whenever game mechanics change. To address this, the researchers tested two alternatives: BERT, which treats the game state as a flexible sequence of data, and GAT, which represents the game board as a graph where tiles are nodes and their relationships are edges. These approaches allow the models to adapt more naturally to varying board sizes and complex, non-linear connections.
How the Models Work
The study compares several architectures:
BERT: Tested in both text-based and board-based formats, these models use self-attention layers to identify dependencies within the game state, allowing for a highly flexible representation of the board.
GAT: These models treat the board as a graph. By using attention mechanisms, they can dynamically weigh the importance of different tiles, even those far apart. The researchers introduced a "GAT2Edges" variant, which simplifies the process by predicting moves directly from the relationships between tiles, reducing computational complexity while maintaining high performance.
Performance and Results
The experiments evaluated the models based on their ability to predict human moves and estimate the difficulty of levels, measured by "Attempts Per Success" (APS). The results showed that the GAT architecture outperformed the CNN baseline in move prediction accuracy. By explicitly modeling the board as a graph, the GAT model successfully captured complex interactions, such as portal connections, that grid-based CNNs cannot represent. While the BERT models offered high flexibility, the GAT approach proved to be a more effective balance for capturing the specific relational topology of the game.
Considerations for Implementation
While the GAT and BERT models offer improved generalization, they come with different trade-offs. The BERT models, particularly the board-based version, require a higher number of parameters and more computational power. The GAT models are more efficient than BERT in terms of floating-point operations, though they still require more inference time than the highly optimized CNN baseline. Ultimately, the study demonstrates that moving toward graph-based representations is a promising path for creating more robust, maintainable, and human-like automated playtesting systems.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!