Human vs Computer Created PuzzlesAre computer created puzzles as good as ones made by humans? It is an interesting question that has been asked many times.
In order to answer it, it seems reasonable to come up with a test to see if a person can distinguish between the two - if a person can tell whether a puzzle comes from a computer or a human then it seems reasonable to say that there is a difference; if they cannot tell then it seems reasonable to say there is no noticeable difference between the two.
So, a good computer created puzzle should be one that a human solver won't know was made by a computer not a human.
Now, for the puzzle itself. It is generally thought and accepted that some puzzles are much harder for a computer to create than others. Clasically, number or logic based puzzles such as sudoku are quite frequently computer made. Can a human tell the difference between the two?
This would be an interesting test, and it requires qualification. First of all, the puzzle written by the human would need to be written by an experienced sudoku setter, and the puzzle set by the computer would need to be from a good piece of software, written by someone who understands the puzzle and how to generate puzzles that mimic how humans solve, dubbed human-solve algorithms.
A human-set puzzle by someone with no detailed knowledge of sudoku, such as the sorts of rules used to solve them, symmetry and perhaps looking for interesting solve paths would be a poor puzzle, and a sudoku from a generator that simply checks every combination would be a poor puzzle. The puzzle must be more than valid - it must be possible for a human to solve using logical rules. In fact this shows some stipulations would be required as to how the puzzle was set - a human or computer could write a puzzle that requires guessing, but since this is not how most people experience playing sudoku, a set of constraints should be set for both human and computer, and no guessing should be one of those requirements.
So the test should screen very good computer generated puzzles and very good human made puzzles, first off.
Whether a test has ever been done with sudoku I'm not sure, but it would certainly be interesting to see if it were done what the result was - and if there was a difference between experienced sudoku solvers and the vast majority - casual newspaper solvers - in distinguishing between human vs computer made puzzles.
Now onto a much tougher challenge - a cryptic crossword. Would a computer be able to set a cryptic crossword, and make it indistinguishable from a human-made puzzle?
Again, it would be a good test to come up with to find out. Certainly it seems that in this case, unlike perhaps sudoku, the computer solver would be limited in its freedom to create a good puzzle.
That is, unless it was extremely clever in a way that may or may not be possible, it would need to work from a preset list of clues for words in its database to have a good chance of coming up with a good puzzle. As far as I'm aware, there is no computer that comes up with good cryptic clues by itself, like a human can. But if there is, it would be fascinating to see it in action!
Working from a preset list of clues, then could the computer create a good cryptic crossword? Well, if the clues in the database are good, then the individual clues in the resulting puzzle would clearly be good too. However, would the puzzle as a whole? Well, often cryptics are themed, and creating a themed cyptic would be more of a challenge. Also since each clue writer tends to have their own 'style' it would be interesting to see how the computer coped with this. Could it have an algorithm to assess clue style so that the resulting puzzle did not appear too discordant? Or would it simply need to have clues tagged by writer and therefore only be able to create a small number of puzzles, unless it had a very large database of clues per writer.
In terms of theming, if the clues were diligently tagged with themes, then it would seem that the computer could create a good puzzle in principle. And if this puzzle passed the test - not many humans could tell it was computer generated - then would it be the case that even the cryptic crossword was not out of reach for a computer generator?
Here a final distinguishing element can be brought to bear - the soft test for computer puzzle creating ability and the hard test. The soft test is one whereby the computer uses knowledge of human clues to create new puzzles, as outlined above simply using a database of pre-written clues.
The hard test is one where the computer simply has access to a list of words, just like the human solver, and has to come up with the clues itself: fully emulate the human cryptic creation process.
Whether there is any program that can do this at present I don't know, but it would be fascinating to see. Certainly it would need a lot of training and some complex algorithms, let's call it cryptic crossword artificial intelligence, to be able to come up with the range of clue types that you might expect to see in a puzzle: clues where the answer is in the word, pun type clues, anagram clues combined with definitions, and a whole host of other tricks you see in cryptic clue setting.
Now, if that program could be made, and could fool humans, then perhaps computers could fairly be said to be able to challenge humans at what seems the hardest of the standard puzzle types for a machine to make.
Does anyone know whether a test has been done to pit the wits of computers against humans in puzzle creation, either for sudoku, crosswords or any other puzzle type? We'd love to hear your thoughts and comments on this.
Date written: 13 Jun 2011
Back to World of Puzzles