The Sudoku problem has many symmetries and permutations to which it is invariant or equivariant. The currently used generators works by sampling existing instances from a fixed database. To improve both the diversity of the data and the generalization capabilities of the agents it could be beneficial to apply data augmentation to the existing instances in the database.
One strategy could be to draw a random data augmentation each time an instance is sampled from the database.
Possible permutations include:
Permute the digits
Rotate the board through multiples of 90 degrees
Reflect the board in any axis (horizontal, vertical, 2 diagonals)
Permute rows within a box (eg any valid permutation of rows (0, 1, 2), of (3, 4, 5), and of (6, 7, 8)), and likewise for columns
Permute groups of rows/columns as boxes (e.g. swap 0&3, 1&4, and 2&5 simultaneously)
The Sudoku problem has many symmetries and permutations to which it is invariant or equivariant. The currently used generators works by sampling existing instances from a fixed database. To improve both the diversity of the data and the generalization capabilities of the agents it could be beneficial to apply data augmentation to the existing instances in the database.
One strategy could be to draw a random data augmentation each time an instance is sampled from the database.
Possible permutations include:
See this for a more comprehensive view.