Thursday, March 28, 2013

[qpenfzsg] Parallelized chess play out

From a given chess position, have the computer play against itself until the end.  Naively this is a time-consuming inherently serial computation, because one cannot start computing the next position until the previous move is known.

Use speculation to achieve parallel computing speedup.  A quick shallow search reveals a good guess what the move will be.  Devilish details and a complicated parallelization pattern, though a very straightforward objective function to optimize: wall clock time.

Motivation is to provide a concrete result for spectators about how an alternate move, or the current position, might play out assuming a high level of play.

Apply distributed analysis to have humans improve the main line.

No comments :