def expand(n):
"""expand the leaf node by adding all its children states"""
if not n.children and not game.terminal_test(n.state):
n.children = {
MCT_Node(state=game.result(n.state, action), parent=n): action
for action in game.actions(n.state)
}
return select(n)
def monte_carlo_tree_search(state, game, N=1000):
def select(n):
"""select a leaf node in the tree"""
if n.children:
return select(max(n.children.keys(), key=ucb))
else:
return n
def expand(n):
"""expand the leaf node by adding all its children states"""
if not n.children and not game.terminal_test(n.state):
n.children = {
MCT_Node(state=game.result(n.state, action), parent=n): action
for action in game.actions(n.state)
}
return select(n)
def simulate(game, state):
"""simulate the utility of current state by random picking a step"""
player = game.to_move(state)
while not game.terminal_test(state):
action = random.choice(list(game.actions(state)))
state = game.result(state, action)
v = game.utility(state, player)
return -v
def backprop(n, utility):
"""passing the utility back to all parent nodes"""
if utility > 0:
n.U += utility
# if utility == 0:
# n.U += 0.5
n.N += 1
if n.parent:
backprop(n.parent, -utility)
root = MCT_Node(state=state)
for _ in range(N):
leaf = select(root)
child = expand(leaf)
result = simulate(game, child.state)
backprop(child, result)
max_state = max(root.children, key=lambda p: p.N)
return root.children.get(max_state)