Algorithm 1 The R-MOTiFS Algorithm

Load dataset and preprocess

Initialize SCALAR, BUDGET //Scaling factor & Number of MCTS simulations (hyper parameters)

function R-MOTiFS (featuresList)

create rootNode

maxReward, bestFeatureSubset ← UCTSEARCH (rootNode)

if maxReward is greater than optimalReward then

optimalReward ← maxReward

optimalFeatureSubset ← bestFeatureSubset

R-MOTiFS (bestFeatureSubset)

else

return (optimalReward, optimalFeatureSubset)

function UCTSEARCH (rootNode)

Initialize maxReward, bestFeatureSubset

while within computational budget do

frontNode ← TREEPOLICY (rootNode)

reward, featureSubset ← DEFAULTPOLICY (frontNode.state)

BACKUP (frontNode, reward)

if reward is greater than maxReward then

maxReward ← reward

bestFeatureSubset ← featureSubset

return (maxReward, bestFeatureSubset)

function TREEPOLICY (node)

while node is non-terminal do

if node not fully expanded then

return EXPAND (node)

else

node ← BESTCHILD (node, SCALAR)

return node

function EXPAND (node)

choose a $\in$ untried actions from A(node.state)

add a newChild with f(node.state, a)

return newChild

function BESTCHILD ($v$, $C$)

$\mathbf{return} \underset{v^{\prime }\in \mathrm{children} \mathrm{of} v}{\max }\max \left(Q_{v^{\prime }}\right)+C\sqrt{\frac{2 \times ln\left(v.visits\right)}{v^{\prime }.visits}}$

function DEFAULTPOLICY (state)

while state is non-terminal do

choose a$\in$ A(state) uniformly at random

state ← f(state, a)

traversestate.path

if ai is equal to fi+1 then

featureSubset ← INCLUDE (fi+1)

reward ← REWARD (featureSubset)

return (reward, featureSubset)

function BACKUP (node, reward)

while node is not null do

node.visits ← node.visits + 1

if reward > node.reward then

node.reward ← reward

node ← node.parent

return