I logged just under 25 hours of active work on this code, which reached 50th on the leaderboard when I retired, about 50 hours into the contest. 22 hours of inactivity later, I was at 113th of the 194 teams which managed to submit scoring plays.
The organisation of the ~450 lines of Haskell code is a bit haphazard,
with some misleading comments. If in doubt, follow the code down from
Main.
The gameplay logic is in GamePlay.hs, where playCommand can be
given a Command to advance one GameState to a new GameState.
The code in AI uses playCommand to explore paths taking the
current unit to all the possible locked positions, then chooses the
"best" position from which to move on to placing the next unit. "Best"
is a heuristic based on the score for that move, plus a bonus for
moves which result in more filled cells lower on the grid. Simple, but
fairly effective.
Haskell was a great choice for this challenge, because the time I spent scratching my head was more than compensated for by the ease of expressing algorithms, and the guarantee of immutability.
In due course I hope to write a full post-mortem of what was a wonderful - if exhausting - experience.
-Steve Purcell
Get in touch if you'd like to work with me at powershop.co.nz -- we're hiring excellent developers.