Feedback and Questions

I recognise that puzzle. I posted an article on it:

http://www.sudokuwiki.org/Arto_Inkala_Sudoku

Relatively famous

It's a mystery where cookies are stored, and different with different browsers.

Do the cookies look like they are doing their job?
There are two on the solver:
1) is automatic and saves the board when you enter a puzzle or 'take step'. That is used to restore the board in case you accidentally close the windows, or just when you come back.
2) there is a manual one - you have to press "save" and "reload" to get that. The board must have changed in order for you to notice a difference when you reload.

Some minor cookies to remember check boxes as well

Yes there is a degree of subjectivity in the ordering of the more difficult puzzles. That's unavoidable and I'd like to offer a solver where the visitor can re-rank them. That's a difficult thing to implement on a webpage already quite complex so I have hesitated. Any re-rank would require recalibration of all the grades - which I do now and again, but it's a big task. So over time the ordering has become fixed. I agree many people would prefer a slightly different order. What I really want to do is branch through all possible orderings of the strategies to find out if there is a super-minimal solve path. Computationally expensive, especially for a multi-user server, but it's where I'm heading. Then a ranking wouldn’t be necessary since it would find the 'best'. Ultimately, the solver is mechanical - it imitates the human only in the contrived and limited sense that some strategies are obviously simpler than others - but it will always fail to imitate how humans solve the puzzles especially with regards to intuition. So there's no easy answer but an interesting topic

Programming the strategies is the fun part, but I did so gradually (and in bursts) over quite some time. Some pseudo-code would be possible but only after explaining some quite complex data structures. I program in C/C++ and make use of a lot of pointers and arrays of pointers which isn't easy to abstract out. The second issue is that the constraints of rows/columns/boxes are also abstracted in the code - so that it is possible to add additional arbitrary constraints to capture other variants like Sudoku X, Jigsaw, Colour Sudoku etc. My Core Engine has to cope with a variety of puzzles and puzzles of different sizes.

Even if a pseudo-code is possible the programmer would never be sure that their implementation was complete and consistent since there is a loss of information in going from code -> pseudo-code. Also pseudo-code prevents the programmer from having a deep understanding of the problem being solved and therefore knowledge of completeness/consistency. So it can actually be harmful, in a sense, they would be writing a user-interface to a black box. Really, the best way is to start from the ground up, and being all your own work, is very satisfying. Expect to go through many incremental versions. I am over a 100 now. The recipes are the descriptions on the website and these contain examples. You can test your progress using sets of difficult puzzles, eg Ruud's top50k. I don’t mind providing those.

Thirdly, the software is commercially sensitive and an asset of the company I work for, as we try and scratch a living on the periphery of a market dominated by a few big players.. ; )

Are you looking at the website? What pages? I don’t have a Nexus but I might be able to suggest something

That's an interesting question. Some people seem to have an amazing eye for spotting chains and seeing patterns in the candidates. For them these very complex advanced strategies are possible with pen and paper - provided that all the candidates have been properly mapped out. For other people they do need an aid - many use the solver to get to a specific bottleneck and print out the board at that stage. That guarantees that the board and candidates are set out correctly and they start looking for links - units where candidates show up just twice. I'm sure they combine this with what-if questions - supposing or removing a number in a cell - but that's using a more trial and error approach. It's still logical in a deterministic way, but my solver avoids deliberate subtraction as I'm interested in pattern based strategies. To that end I'd say that the bulk of the 'extreme' strategies are veering towards the academic rather than the practical - although as I said, some people I don't doubt have used them without a hint or a solver.

What I'm doing with the book and the solver/website is to collect all the known valid ways of using pattern based (or 'state deterministic') ways of logically removing candidates and some are pretty esoteric to say the least. A computer is ideal for searching for them - as a coder I've gotten great satisfaction from devising optimal ways for catching all instances of a strategy using the least amount of time/memory - but that is a different activity from a pen/paper solver trying to solve a specific puzzle.

As you investigate a board and map out all the links in chains it's not really possible to know in advance which strategy will be most useful next. The solver gives the illusion of knowing but only because it is so fast. If it picks Medusa Rule 3, it does so only because it has exhaustively searched for and failed to find any instance of a simpler strategy. We humans can't do exhaustive searches, we have to use hunches and intuition. We will probably miss a simpler path when we use a complex one - but there are many, many paths (usually).

I would say, though, that the most fruitful area is the family of chaining strategies - you can build up a series of links between bi-value and bi-location candidates - these are the building blocks of all those strategies. Merely having an awareness of the tricks that each strategy will, in time, seep into your image of the board state and you'll get the Eureka moment.

I'd like to hear from pen and paper solvers who have tips and 'rules of thumb' they use. I'm sure there are a lot of interesting approaches. Some people swear by using dots in a 3x3 grid rather than small numbers, for example.

Thanks for the kind feedback, most appreciated! And very glad you enjoy the site

Been nine years, but slowly and in bursts! Had many excellent ideas and contributions to work with. Glad you find it useful

Doh, correct. Will change in next update, thx!

That does look like a very difficult puzzle. So hard to compare. I have some ideas here but I will need to adapt my program for this to work with Sudoku X.