Kurodoko solver

March 28, 2021

I wrote a Kurodoko solving program! I’m rather pleased with it. Kurodoko is a puzzle genre invented by Nikoli.


I have twice attempted to write code to solve Slitherlink puzzles—my favourite Nikoli puzzle type—but never gotten far. This time, I had two advantages:

  1. Kurodoko is a much simpler puzzle type than Slitherlink. In Slitherlink, you have to worry about cells and edges; in Kurodoko, we only need to think about cells, so the code feels simpler. Also, the ‘basic starting deductions’ in Slitherlink feel more complex than in Kurodoko.

  2. I was recently taught the practice of Test-Driven Development! In TDD, you do not write new code until you’ve written a test for it. This promotes complete test coverage, and, maybe more importantly for me, it encourages you to break the problem down into truly bite-sized requirements.

With these advantages and a fresh dose of enthusiasm (the result of cracking open a new Nikoli puzzle book), I was able to knock out a working Kurodoko solver in one weekend! Version 0.1 was only able to solve simple puzzles since it did not have the ability to search branches of the solution space, but I returned to the project on later weekends: v0.2 was able to make more deductions and could solve the example puzzle by Cross-Plus-A, and v0.3 could solve the example puzzle on Wikipedia and recognize grids that are infeasible or unconstrained. The fact that I was able to continue development where I left off after weeks away, and make brisk progress, has impressed upon me the value of TDD.

The code can be viewed on the Github page: Kurodoko Solver.

Why do this?

Two main reasons:

  • It seemed fun! 🤗
  • An automatic solver is a useful aid to constructing puzzles, since it allows one to check whether a grid is solvable, unconstrained or infeasible. This is why, when writing the solver, I wanted to restrict to a set of logical deductions (and shallow search depth) that I knew human solvers were capable of.

What’s next?

I have a few TODOs:

  • Print grid images to PDF
  • Test it out on a larger set of puzzles
  • Make the deduction structure clearer and more efficient
  • Genericise the project to handle similar puzzles? (Kurodoko seems almost isomorphic to the Corral type, and modelling Corral is a clear stepping stone to modelling Slitherlink.)
  • Make a web app where users can design grids with an interface and run the solver on them to see if they’re solvable?