Perception-based constraint solving for sudoku images

Area of Science:

• Artificial Intelligence
• Computer Vision
• Constraint Satisfaction

Background:

• Perception-based constraint solving integrates user-provided images into problem specifications.
• Neural networks interpret images to extract data, such as Sudoku grid values.
• Existing methods struggle with noisy or erroneous image data.

Purpose of the Study:

• Investigate a hybrid machine learning and constraint solving approach for joint inference.
• Evaluate the impact of classifier calibration on joint inference performance.
• Develop a robust framework to identify and correct user errors in perception-based constraint solving.

Main Methods:

• A hybrid model combining machine learning for perception and constraint solving for reasoning.
• Joint inference predicts blank cells and solves the puzzle simultaneously.
• Classifier calibration techniques are applied to improve prediction accuracy.
• Constraint satisfaction identifies inconsistencies arising from user errors.

Main Results:

• Joint inference successfully corrects errors made by the perception classifier.
• Classifier calibration enhances solution quality across various datasets.
• The framework demonstrates robustness in handling user-written errors and distinguishing input types.

Conclusions:

• Hybrid perception-based constraint solving offers a powerful approach for complex problems.
• Classifier calibration is crucial for improving the accuracy of joint inference.
• The developed system effectively handles real-world imperfections, including user errors, for robust problem-solving.