Friday, March 30, 2012

Optimal Janowski cube life index revisited

After fixing my implementation of Janowski's doubling strategy I tried again to find the optimal cube life index (assuming we just use a single constant index).

As before, I found the optimal cube life index is 0.7.

To test this I ran doubling strategies with different cube life indexes against one with a fixed index of 0.7 to see if any other index would outperform. For indexes near 0.7 I ran 1M self-play games, using Player 3.3 for checker play. For indexes a bit further out I used 100k self-play games.

Here are the results:

Two interesting points to note. First is that the probability of win continues to increase to x=0.95 even though the average score decreases, due to the asymmetric market windows of the two players. Second is that there is a fairly dramatic falloff in performance (both in score and probability of win) for x>0.95.

No comments:

Post a Comment