-
Notifications
You must be signed in to change notification settings - Fork 1
/
mcts.py
162 lines (140 loc) · 5.56 KB
/
mcts.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
import game as emulator
import math
import numpy as np
the_emulator = emulator.Environment()
class Node:
def __init__(self, parent, layer, action, value=1, node_reward=0, is_end_node=False):
self.parent = parent
self.value = value
self.action = action
self.node_reward = node_reward
self.is_end_node = is_end_node
self.play_count = 1
self.child_nodes = []
self.layer = layer
def get_child_nodes(self):
return self.child_nodes
def add_child_node(self, action, value, node_reward, is_end_node):
new_node = Node(self, self.layer + 1, action, value=value, node_reward=node_reward, is_end_node=is_end_node)
self.child_nodes.append(new_node)
return new_node
def get_best_child_according_to_ucb(self, mcts_obj):
if len(self.child_nodes) > 0:
max_node = self.child_nodes[0]
max_val = max_node.value
for child in self.child_nodes:
if child.is_end_node:
return child
my_val = child.value + math.sqrt((2 * math.log(mcts_obj.n)) / child.play_count)
# print(mcts_obj.n, my_val, child)
if my_val > max_val:
max_val = my_val
max_node = child
return max_node
def get_most_played_child(self):
if len(self.child_nodes) > 0:
most_node = self.child_nodes[0]
most_plays = most_node.play_count
for child in self.child_nodes:
if child.is_end_node:
return child
my_play_count = child.play_count
if my_play_count > most_plays:
most_plays = my_play_count
most_node = child
return most_node
def has_childs(self):
return len(self.child_nodes) > 0
def has_all_childs(self):
return len(self.child_nodes) >= the_emulator.action_size
def get_parent(self):
return self.parent
def get_child_count(self):
return len(self.child_nodes)
def has_parent(self):
return self.parent is not None
def __str__(self):
return f"layer: {str(self.layer)}, action: {emulator.print_action(self.action)}, value: {str(self.value)}, " \
f"plays: {str(self.play_count)}"
class MCTS:
def __init__(self):
self.tree = Node(None, 1, None)
self.state = None
self.agent = None
self.n = 0
self.cur_node = self.tree
def do_mcts(self, first_state, agent):
self.state = first_state
self.agent = agent
for i in range(20000):
if (i + 1) % 20 == 0:
print((i * 100) / 20000, "% mcts done")
leaf_node, leaf_state, prev_reward = self.select()
if leaf_node.is_end_node:
new_leaf_node = leaf_node
else:
new_leaf_node = self.expand(leaf_node, leaf_state, prev_reward)
self.backpropagate(new_leaf_node)
cur_node = self.tree
actions = []
total_rew = 0
while cur_node.has_childs():
cur_node = cur_node.get_most_played_child()
print(cur_node.action, cur_node.value, cur_node.node_reward)
actions.append(cur_node.action)
total_rew += cur_node.node_reward
if cur_node.is_end_node:
return total_rew, actions
return total_rew, actions
def simulate(self, state):
total_reward = 0
steps = 0
s = state
# print("new sim")
while steps < self.agent.handler.hp.MAX_STEPS:
steps += 1
prediction = self.agent.brain.predictOne(the_emulator.get_flattened_reduced_state(s))
a = np.argmax(prediction)
# print(emulator.print_action(a))
s_, r, done, info = the_emulator.step(a, flattened=False, state=s, hyperparams=self.agent.handler.hp)
s = s_
total_reward += r
if done:
break
self.n += 1
return total_reward
def backpropagate(self, leaf_node):
cur_node = leaf_node
while cur_node.has_parent():
cur_node.play_count += 1
cur_node = cur_node.parent
childs = cur_node.get_child_nodes()
cumulative = 0
for child in childs:
cumulative += child.value
cur_node.value = cumulative / cur_node.get_child_count()
def expand(self, node, state, prev_reward):
prediction = self.agent.brain.predictOne(the_emulator.get_flattened_reduced_state(state))
child_c = node.get_child_count()
for i in range(child_c):
prediction[np.argmax(prediction)] = -999999
a = np.argmax(prediction)
s_, r, d, _ = the_emulator.step(a, flattened=False, state=state, hyperparams=self.agent.handler.hp)
if not d:
rew = self.simulate(s_)
else:
rew = r
return node.add_child_node(action=a, value=rew + prev_reward, node_reward=r, is_end_node=d)
def select(self):
prev_reward = 0
cur_node = self.tree
s = np.array(self.state)
while cur_node.has_all_childs():
cur_node = cur_node.get_best_child_according_to_ucb(self)
prev_reward += cur_node.node_reward
s_, r, d, _ = the_emulator.step(cur_node.action, flattened=False, state=s, hyperparams=self.agent.handler.hp)
s = s_
if cur_node.is_end_node:
self.n += 1
break
return cur_node, s, prev_reward