Learning to Score Final Positions in the Game of Go

Key insights

6 takeaways

01
Machine learning models achieve 98.9% accuracy on 9x9 Go boards and 99.4% on 19x19 boards by classifying the life-and-death status of stone blocks.See it on page 1
02
The system utilizes a recursive partition neural network trained on human game records to replace static knowledge bases with adaptive, learned representations.See it on page 1
03
Models trained exclusively on 9x9 board configurations demonstrate high performance on 19x19 boards, proving the methodology is effective at generalizing across different board sizes without additional training.See it on page 14
04
By shifting from raw board data to a specialized, feature-based representation, the system effectively resolves the complexities of final game state evaluation.See it on page 15
05
The integration of a recursive classification framework allows the system to refine its assessments through iterative analysis, enhancing overall performance.See it on page 13
06
This automated scoring foundation provides a scalable framework for future AI development, particularly for evaluating the higher strategic complexity of mid-game positions.See it on page 1

Summary

Machine learning systems can achieve near-perfect accuracy in determining the final scoring of Go games by classifying the life-and-death status of stone blocks. By moving away from raw board data toward a specialized, feature-based representation, these models effectively resolve the complexities of board state evaluation. The integration of a recursive classification framework further enhances performance, allowing the system to refine its assessments through iterative analysis.

The methodology relies on training neural network classifiers, specifically recursive partition neural networks, on extensive human game records. This approach demonstrates remarkable robustness, as models trained exclusively on 9x9 board configurations maintain high performance when applied to larger 19x19 boards. The system achieves a 98.9% accuracy rate on 9x9 grids and an even higher 99.4% accuracy on 19x19 boards, confirming that static evaluation is a highly reliable method for resolving final game states.

These findings establish a significant benchmark for computational game theory, proving that static knowledge bases can be successfully replaced by adaptive, learned representations. By successfully automating the scoring process, this research provides a scalable foundation for future developments in artificial intelligence, particularly for evaluating non-final, mid-game positions where strategic complexity is significantly higher. The ability to generalize across different board sizes without additional training highlights the efficiency and potential of this feature-based approach in broader game-playing applications.