Efficient Selectivity and Backup
Operators in Monte-Carlo Tree Search
Presented by Melvin Zhang
R´emi Coulom, Computer and Games 2006
http://rusgo.org/crazystone/
http://www.wired.com/2014/05/the-world-of-computer-go/
Selection
Principle: Move that look best should be searched deeper, and
bad moves should be searched less
Selection
Principle: Move that look best should be searched deeper, and
bad moves should be searched less
Crazy Stone: Each move i is selected with probability
proportional to
exp −2.4
µ∗ − µi
2(σ2
∗ + σ2
i )
+ i
Backup/Update
Update the statistics at internal nodes base on result of
simulation. Affects subsequent selection.
Backup/Update
Update the statistics at internal nodes base on result of
simulation. Affects subsequent selection.
Crazy Stone: Uses “Mix” operator, linear combination of mean
and robust max
mean
average of child nodes
max
max of child nodes
robust max
child node with the most simulation
http://peepo.com/go.html
https://code.google.com/p/magarena/

Efficient Selectivity and Backup Operators in Monte-Carlo Tree Search