Single Comment

csipgs Chess. Design and buy new chess pieces during play. (8x8, Cells: 64) [All Comments] [Add Comment or Rating]

Jeremy wrote on Fri, Mar 25, 2011 03:43 AM UTC:

'Computer programs cannot play this game competently.'

What makes you say that?  If I were going to design an AI for this game,
I'd probably:
1.  Start with something like a FIDE Chess AI
2.  Modify the heuristic for the value of a board position.  Estimate piece
value based on cost, reserve pieces about a tempo less, bank may be
somewhat nonlinear but this doesn't seem scary.
3.  Create a representative library of about 20-50 useful pieces, including
at least 1 using each atom.
4.  For lookahead, you know the pieces purchasable on the current turn;
assume pieces purchasable on future turns are that player's current
designs plus entire library (limited by bank).
5.  Collect some stats from the lookahead to predict the designs you'll
need in the immediate future (AI never builds designs except those in
library).

Branching may be higher than FIDE, enough to weaken the AI, but I would
think it could still play respectably.  You can sandbag the AI if you
invent a piece that is significantly better for the current position than
any piece in its library, but that seems hard, and AI starts considering it
as soon as you design it, which is at least 2 turns before it can attack. 
Am I missing something?