a1k0n wrote:The difference in endgame moves after separation is roughly equal to K1 * (player 1 fillable squares - player 2 fillable squares) + K2 * (player 1 sum of edges - player 2 sum of edges) + K3 * (irrelevant factor). [...] I found K1 to be about 0.050, K2 about 0.150, and K3 less than 0.009. Hence I consider the third factor pretty much irrelevant.
Wow, thanks for sharing this. Your earlier brief summary made me wonder how you topped the standings using only ideas that had been openly discussed. Was your implementation simply blindingly fast? No, you (also?) had a better eval! Like you I'm surprised at the empirical importance of edges rather than vertexes, but hey, whatever works. Those factors penalize holes in the map that our eval didn't penalize at all, and holes are what always stop us from reaching the upper bound. Great idea to fit the parameters to actual game results.
On a totally unrelated note, I'm rooting for you because I think it would be hilarious if a Yahoo won the Google AI challenge.