@ 2011-02-28 10:37 PM (#3625 - in reply to #3491) (#3625) Top
Country : United States
MellowMelon posted @ 2011-02-28 10:37 PM
Thank you for all the comments. I'll be keeping them in mind for any puzzles/competitions I write in the future.
Some notes regarding difficulty: The reason I feel a need to apologize about the timing is that I do think the best solvers prefer finishing. For one thing, there's an element of luck introduced when it comes down to which puzzles a solver decided to do, and I think total time is a more consistent separator. Also, from this post by motris regarding the most recent WPC, which was in my thoughts while calibrating the test:
Most of these comments are not too surprising in hindsight, as my post mortem shows, but the Liar Slitherlink baby comments are. The first step, as the solution PDF notes, is that the center two rows have clues only in two columns, so no other clues in those columns can be liars. It is a bit of a different style from the usual ways of breaking in, but I seriously thought this was going to be easier to find than searching for configurations of impossible clues in the whole grid, especially since it was so localized. There was not much indication in testsolvers' times that it would be much tougher to see.
The types being outlandish was definitely an issue. I'll make sure any future contests have, at the least, a less wordy set of rules.
For answer extraction, I considered several different ones for Castle Wall. Re consistency: the Double Back scheme is terrible for Castle Wall because figuring out the general path of the loop is often very easy thanks to the colors, so it's more of a question of whether a loop uses a square or not. That's starting to become possible to brute force. I'm still not sure what I should have used for it. The row/column mechanisms seemed to be the best ones (no counting involved), so I possibly should have just asked for used/unused in two rows or columns. That was considered before the test and thought not to give enough information. Was I right? I still don't really know.
Anyways, I knew from the start that puzzle quality would likely be the main thing I got right in doing this, so I'm glad to hear the positive comments about it. Perhaps one reason the babies were all too hard is that I've always had trouble making a very good puzzle of low difficulty, and I wanted every puzzle on this test to be "very good". I'll be trying to figure this out in the near future.
Some notes regarding difficulty: The reason I feel a need to apologize about the timing is that I do think the best solvers prefer finishing. For one thing, there's an element of luck introduced when it comes down to which puzzles a solver decided to do, and I think total time is a more consistent separator. Also, from this post by motris regarding the most recent WPC, which was in my thoughts while calibrating the test:
The rounds probably also suffered in general from having too many puzzles or too little time, and I was hurt by a strategy of trying to go through a round as I would to finish it instead of just going for 50% of the points as fast as I could. Round 9, where Ulrich topped with 60% of the total score was the big round I lost momentum and fell from essentially tied to 80 points back, because I went through puzzles in an order planned to finish many more than I did. Only three solvers finished rounds (1 in round 1, me in round 2, and 1 in Polyomino) which seems very low at a WPC. I suppose we only had 7-10 at the WSC but there I think we errored ourselves by using my time as the mean time too frequently and not making some of our puzzles easier. It is ok for solvers to finish rounds.
Most of these comments are not too surprising in hindsight, as my post mortem shows, but the Liar Slitherlink baby comments are. The first step, as the solution PDF notes, is that the center two rows have clues only in two columns, so no other clues in those columns can be liars. It is a bit of a different style from the usual ways of breaking in, but I seriously thought this was going to be easier to find than searching for configurations of impossible clues in the whole grid, especially since it was so localized. There was not much indication in testsolvers' times that it would be much tougher to see.
The types being outlandish was definitely an issue. I'll make sure any future contests have, at the least, a less wordy set of rules.
For answer extraction, I considered several different ones for Castle Wall. Re consistency: the Double Back scheme is terrible for Castle Wall because figuring out the general path of the loop is often very easy thanks to the colors, so it's more of a question of whether a loop uses a square or not. That's starting to become possible to brute force. I'm still not sure what I should have used for it. The row/column mechanisms seemed to be the best ones (no counting involved), so I possibly should have just asked for used/unused in two rows or columns. That was considered before the test and thought not to give enough information. Was I right? I still don't really know.
Anyways, I knew from the start that puzzle quality would likely be the main thing I got right in doing this, so I'm glad to hear the positive comments about it. Perhaps one reason the babies were all too hard is that I've always had trouble making a very good puzzle of low difficulty, and I wanted every puzzle on this test to be "very good". I'll be trying to figure this out in the near future.