How Good Are Computer-Generated Puzzles?

The background: I'm a Maleska-era dinosaur who's thinking of getting back into construction.  I believe that the general quality of Shortzian puzzles is a fair bit higher than the puzzles I (and I suspect my Maleskan peers) made in the early 1990s.  I have a defensive suspicion that part of the new-era improvement comes from constructors using software that we oldsters didn't.   But how good is that software, anyway? 

Rex's noting today that his crossword-constructing software might have once spat out SO RARE, which he would have disdained as fill (I wouldn't--what is so rare as a day in June?) pushed me over the edge to find out...

At this very moment, Crossword Compiler is working away on a very pretty demo grid for me--a much prettier one than I ever attempted in my constructing days, with an impressively low (for me) number of words (66) and black squares (25).  

I gave the software two words--CALIFORNIA and CALLIPYGIAN--as an implicit theme.   I'm very impressed by its "thought process" as shown on screen.  It simulates a human constructor nicely by putting in some good quality long fill--OPPORTUNISM, CORONER, and YARDARM--and then trying to build out from there. 

After 28 minutes, the program hasn't gotten very far--it's filled in a reasonable SW with the mediocre but hardly awful ASSAYER and RETINUE as core elements, but then keeps Xing the SW out as it tries to do the fill in the SE.   Words flash by so quickly it's hard for me to follow them on the screen--nice ones like CASANOVA and CILANTRO, okay ones like SENATORS, and weak ones like DESALTER are ones I've seen for the eight-letter crosser for the second A of YARDARM.  

I'm wondering if the program is cycling at this point and simply won't announce failure even though it's out of possibilities.   On the other hand, it told me it couldn't do the grid when I put in JAMBALAYA as 1D along with CALIFORNIA and CALLIPYGIAN, so I still have hope--or should I say fear--that the program will by the end of the day come up with a good fill for the grid, something I'm confident I couldn't do in the same amount of time.

Even if Crossword Compiler in the end can't do the grid with the CALI-CALLI fill, I'll remain highly impressed by its simulation of a human constructor.   And yes, I will use it or something like it as an aid if I do return to the constructing lists...will have to figure out appropriate boundaries between myself and the software if I do that. 

Fascinating how CWPs generate endless ethical issues for solvers, editors, and constructors...

 [1 hr. update--Compiler has abandoned OPPORTUNISM and YARDARM for IMPORTATION and YARDMAN...oh well...at least it tried the prettier long fill first, which counts in its favor]

[2 hr. update--now it's trying DECOLLETAGE with YARDAGE...I like, but it's also using FAECAL, which is kind of edgy even for moi]

[140 min.--now it's EMBOLDENING w/ YULE LOG...fine, but it's still stuck on FAECAL...oh well...am guessing the program doesn't have a whole lot of multi-word phrases to draw on beyond highly salient ones like YULE LOG...suspect that's an area where John Henry still beats the machine]

[150 min.--LABORSAVING(!) and YARNING, with ALLEGRETTO the first crosser for the G of CALLIPYGIAN I've seen that's lasted for more than a semi-demi-sec]

Gotta do some work today...will check on Compiler's progress or lack thereof in two hours when I break for lunch and then again at the end of the day...regardless of the outcome, I find the simulation of human construction by the software to be highly fascinating and not a little disturbing..ah, the wonders of 2013]

[270 minutes...lunch break for me!...but not for Compiler...it's still worrying over the same part of the puzz...now it's on REPOSSESSED, Y-SHAPED, and FOEMAN...it ain't fast but it never gets tired]

[324 minutes...gotta get back to work...CC still slogging away on REPOSSESSED and Y-SHAPED, with an okay SW and the elusive SE still up in the air]

[7:33 p.m...at home...left my laptop at work in rush to catch train...will check in the am to see what's up]

[7:34...semi-bright idea...can run the same grid with another Compiler demo on this laptop and see if the results are exactly the same or whether there's a random element in the program]

[7:43...okay, done...so far it looks like CC2 is doing the same things as CC1 w/ the same grid...same starting pattern of OPPORTUNISM, CORONER, and YARDARM]

[back at work...CC1 slept w/ my laptop at night...it's only up to 495 minutes...it's stuck in the same old same old...now it's trying STOOL PIGEON, YUCATAN, and COLONEL as the anchors]

[lunch time!...758 minutes in...now the anchors are TAM-O-SHANTER and YIELDER]

[the tireless CC is now at 935 minutes...it's using ECHO SOUNDER and YIELDER as the anchors, and is playing around some with the SE and NW...also saw HANG GLIDER stay onscreen for a while in the NE...not sure whether that's progress or not]

[experiment terminated semi-accidentally after 1060 minutes...a discouraging REMIGRATED was in the NE along with ECHO SOUNDER and YIELDER]

[insta-fill accomplished with a 76 word 36 block grid w/ CALLIPYGIAN, CALIFORNIA GURLS, and GOLDEN COAST as theme answers...not pretty though...RELAPSER and TEG were a few of the lowlights]

Comments

[loren smith]

I'm a fledgling constructor and use Crossword Compiler. I've never been happy at all with the Autofilled puzzles. I usually check to see if it's doable and then go it alone. Sometimes it's helpful to see what the computer wants for a longish stack, but that's about it for me!

[Wayne]

Thanks, Loren! Am not seeing yr comment here, but got it in an email...think yr point on using software for limited purposes like hard stacks makes sense.