Summary of the algorithm

Given a state $s(0)$ select a move $m(0)$ by

1: Patterns 
$P=\{\{\}\}$, # of play out $k=0$ 


2: Play-out depth $j=0$ 


3: Move 
$m(j)=\arg \max_{m} h_h(m \mid p)+$noise 


4: Make move $m(j)$ in state $s(j)$ to get new $s(j+1)$ 


5: Increase $j$ by one 


6: If $s(j)$ not finished, loop to 3 


7: Increase $k$ by one 


8: Save moves $m(\cdot)$ and the result of $s(j)$ in $c_{k}$ 


9: Loop to 2 for some time 


10: Add new patterns to $P$ 


11: Loop to 2 for some time 


12: 
$m(0) = \arg \min_{m_1} \max_{m_0} h_h(m_0\mid \{m_1\})$

On line 3, the moves of the pattern $p \in P$ and $p \subset \{m(0),\dots,m(i-1)\}$ and there must be no $q \in P$ such that $p \subset q \subset \{m(0),\dots,m(i-1)\}$. On line 10, the candidate patterns are all patterns in $P$ extended by one move, and they are accepted if the particular move has been selected by that pattern more than some threshold number of times on line 3.