A Sudoku solver that uses human strategies.

• Naked singles, pairs, triples, and quads
• Hidden singles, pairs, triples, and quads
• Unit intersections (a.k.a. pointing pairs and box+line reductions)
• X-wings, swordfish, and jellyfish
• Y-wings (a.k.a. XY-wings) and XYZ-wings
• 3D Medusas (a generalization of simple coloring)
• Dual Medusas (3D Medusas that start with a bi-location candidate in a unit instead of a bi-value cell)
• Bi-value cell forcing chains
• Dual unit forcing chains
• Nishio forcing chains
• Anti-Nishio forcing chains (Nishio forcing chains that start with a candidate digit off instead of on)
• Subset exclusion (incomplete)
• Guessing individual cells (with the -g flag enabled)

These strategies are insufficient to solve all boards without guessing. My intention is to automate the strategies that I fully understand, since at that point doing them myself is just busy work.

So far it can solve 99.5% of these 49,151 17-hint Sudoku (only 266 are unsolvable). With the puzzles chosen by Peter Norvig, it can solve all 50 of the easy ones, 64 of the 95 hard ones, and 9 of the 11 hardest ones. However, all of the AI Sudoku Top 10 are unsolvable. (With guessing, all boards are solvable.)

• ./sudoku.py BOARD
  Solves the given board and shows steps.
  e.g. ./sudoku.py 000000001000000020000003000000040500006000300007810000010020004030000070950000000
• ./sudoku.py -f FILE
  Solves each board in the given file and outputs a TSV summary.
  e.g. stdbuf -oL ./sudoku.py -f boards.txt > solutions.tsv
• ./sudoku.py -q BOARD
  Solves the given board without printing anything. Useful for measuring performance.
  e.g. time ./sudoku.py -q 000000001000000020000003000000040500006000300007810000010020004030000070950000000
• ./sudoku.py -g BOARD or ./sudoku.py -g -f FILE
  Solves the given board or file with guessing enabled.
  e.g. ./sudoku.py -g 000000001000000020003004000000003500010060000720000080000108000000720000900000600