Hi,
from
recent posts in computer-go my guess is:
the strongest
programs will use DCNN and
https://www.conftool.net/acg2015/index.php/Graf-Adaptive_Playouts_in_Monte_Carlo_Tree_Search_with_Policy_Gradient_Reinforcement_Learning-113.pdf?page=downloadPaper&filename=Graf-Adaptive_Playouts_in_Monte_Carlo_Tree_Search_with_Policy_Gradient_Reinforcement_Learning-113.pdf&form_id=113&form_version=final
soon.
Both
approaches seem to be quite independent and each giving more than
100ELO. Abakus already bet zen with adaptive playouts and Aya reports
good success with DCNN.
Therefore I prepare for adaptive
playouts. Looking into this showed, that we are really 5 years behind
with playout policy. I did not recognize how different other playouts
were:(
I prepared to use real liberty data structures (did not
harm performance)
I prepared crazy stone playout style (we had
mogo before with modifications into this direction)
I prepared the
first 7 features used in Graf paper, but did not debug yet.
At
the moment I am not sure, if it will ever be possible to nicely pull
from my fork to Francois repository:(