Run python main.py in the terminal and a small 9x9 Sudoku will be generated then solved using a backtracking algorithm.
sudoku.fill()to fill an empty grid with numbers that follow the rules of Sudoku.sudoku.delete()to randomly delete numbers out of the grid to turn it into a puzzle.sudoku.show()to display the sudoku visually using matplotlib just like in the image above.sudoku.save_text()to save a sudoku as a text file.
solver.solve_sudoku()to find the solution for a given Sudoku.
See Kyubyong's Sudoku implementation. It proposes a neural network solver for Sudoku tasks. There is plenty of other code and research on the subject as well. I have a few related questions:
1st question: What variation of Sudoku would require a neural network in order to be solvable in a reasonable time? Scaled up size and more numbers?
2nd question: Can an AI learn the rules from Sudoku and execute them systematically
- without any instructions in form of rules but
- only given the smallest set of possible examples?
This problem is similar to the Abstraction and Reasoning Challenge yet more focused on chains of mathematical reasoning which truly requires code synthesis rather than inferring rules based on core-knowledge and only implementing them manually a few times.
It is also similar to Alpha Geometry as the goal is to find the right chain of operations to solve/complete the puzzle but with the added challenge of understanding the rules of the game given examples in the first place.
I believe it is important to
- A) minimize the amount of examples from which the rules of the game or program are inferred, and
- B) to maximize the amount of steps an AI has to execute and reason through systematically in order to solve a new puzzle that abides to the same rules.