I was just looking over the code for gnugo, I don't know if its representative; but there are tons of parameters (values of given moves primarily, but also various fiddly things buried elsewhere) which seem arbitrary. Why not apply a learning approach (eg self-play to find the exact best values of these? Josh