-
Notifications
You must be signed in to change notification settings - Fork 3
/
PPOFP_test.py
149 lines (130 loc) · 6.12 KB
/
PPOFP_test.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
import numpy as np
import tensorflow as tf
import pickle
import copy
from XFP import XFP, LeducRLEnv
from PPOFP import PPO
FLAGS = tf.app.flags.FLAGS
# Experiment settings
tf.app.flags.DEFINE_integer('seed', 100, 'random seed')
tf.app.flags.DEFINE_integer('num_cpu', 8, 'num of cpus')
tf.app.flags.DEFINE_integer('card_num', 6, 'card numbers in leduc')
tf.app.flags.DEFINE_integer('data_len', 3*10000, 'data set length')
tf.app.flags.DEFINE_integer('epochs', 4, 'training epochs')
tf.app.flags.DEFINE_integer('batch', 256, 'batch size')
tf.app.flags.DEFINE_float('ppo_clip', 0.2, 'PPO clip parameter')
tf.app.flags.DEFINE_float('encoeff', 0.0, 'entropy bonus')
tf.app.flags.DEFINE_float('lr', 0.005, 'learning rate')
tf.app.flags.DEFINE_integer('fsp_iter', 100, 'fictitious play iterations')
tf.app.flags.DEFINE_bool("use_gpu", False, "run on GPU")
xfp = XFP(card_num=FLAGS.card_num, seed=FLAGS.seed)
def evaluate_exploitabality(p1, p2):
xfp.opponent_policy_p2 = p2
xfp.compute_p1_best_response()
br1_op = XFP.convert_q_s_a2greedy_policy(xfp.q_value1_final)
xfp.opponent_policy_p1 = p1
xfp.compute_p2_best_response()
br2_op = XFP.convert_q_s_a2greedy_policy(xfp.q_value2_final)
# compute optimal best response payoff
realization_br1_op = XFP.compute_realization(br1_op, p2)
realization_br2_op = XFP.compute_realization(p1, br2_op)
e = [XFP.compute_payoff_given_realization(realization_br1_op)[0],
XFP.compute_payoff_given_realization(realization_br2_op)[1]]
exploitability = (e[0] + e[1]) / 2.0
xfp.finish()
return exploitability, e
def evaluatate_best_response(iteration, p1, p2, br1, br2):
realization_old = XFP.compute_realization(p1, p2)
realization_rl_br1 = XFP.compute_realization(br1, p2)
realization_rl_br2 = XFP.compute_realization(p1, br2)
mix_realization_p1 = XFP.mix_realization(realization_rl_br1, realization_old, 0.1)
mix_realization_p2 = XFP.mix_realization(realization_rl_br2, realization_old, 0.1)
new_p1, _ = XFP.compute_realization2policy(mix_realization_p1)
_, new_p2 = XFP.compute_realization2policy(mix_realization_p2)
exploitability, _ = evaluate_exploitabality(new_p1, new_p2)
return exploitability, new_p1, new_p2
def train():
agent = PPO(FLAGS)
p1, p2 = xfp.get_uniform_policy()
iteration = 0
while iteration < FLAGS.fsp_iter:
print '---------------Fictitious Self-play Iteration', iteration, '------------------'
exploitability, e = evaluate_exploitabality(p1, p2)
print "optimal best response policy for p1 achieves {:.4f}\n" \
"optimal best response policy for p2 achieves {:.4f}, exploitability is {:.4f}".format(
e[0], e[1], exploitability)
agent.learn(10, [p1, p2])
br_policy, _ = agent.compute_self_policy()
mixture_e, p1_new, p2_new = evaluatate_best_response(iteration, p1, p2, br_policy[0], br_policy[1])
print mixture_e
iteration += 1
p1 = p1_new
p2 = p2_new
def some_tests():
flags = copy.deepcopy(FLAGS)
flags.batch = 4
flags.epochs = 1
agent = PPO(flags)
# test compute policy
agent.compute_self_policy()
for key, item in agent.explicit_policy[0].iteritems():
state_card = np.zeros([agent.env.card_num], np.int32)
card, state = xfp.get_card_state_state(key)
state_card[int(card[0])] = 1
if len(card) == 2:
state_card[int(agent.env.card_num) / 2 + int(card[1])] = 1
state_history = agent.env.history_string2vector[state]
res = agent.sess.run(agent.ops[0]['softmax'],
feed_dict={agent.ops[0]['state_card_inf']: [state_card],
agent.ops[0]['state_history_inf']: [state_history]})
assert res[0][0] == item[0] and res[0][1] == item[1]
for key, item in agent.explicit_policy[1].iteritems():
state_card = np.zeros([agent.env.card_num], np.int32)
card, state = xfp.get_card_state_state(key)
if len(card) == 2:
state_card[int(card[1])] = 1
state_card[int(agent.env.card_num) / 2 + int(card[0])] = 1
else:
state_card[int(card[0])] = 1
state_history = agent.env.history_string2vector[state]
res = agent.sess.run(agent.ops[1]['softmax'],
feed_dict={agent.ops[1]['state_card_inf']: [state_card],
agent.ops[1]['state_history_inf']: [state_history]})
assert res[0][0] == item[0] and res[0][1] == item[1]
# test load parameters
parameters = agent.get_parameters()
agent.sess.run(agent.init)
agent.load_parameters(parameters)
for i, j in zip(parameters, agent.get_parameters()):
assert np.allclose(i, j)
# test collect_sample and data pipeline
agent.opponent_policy = agent.explicit_policy
agent.collect_samples(100)
# print agent.top
data = [{}, {}]
for i in agent.data_returns[0, :agent.top[0]]:
data[0][i] = data[0][i] + 1 if i in data[0] else 1
for i in agent.data_returns[1, :agent.top[1]]:
data[1][i] = data[1][i] + 1 if i in data[1] else 1
data2 = [{}, {}]
for i in range(2):
agent.sess.run(agent.ops[i]['iterator'].initializer,
feed_dict={agent.ops[i]['state_history']: agent.data_states_history[i, :agent.top[i]],
agent.ops[i]['state_card']: agent.data_states_card[i, :agent.top[i]],
agent.ops[i]['action']: agent.data_actions[i, :agent.top[i]],
agent.ops[i]['advantage']: agent.data_advant[i, :agent.top[i]],
agent.ops[i]['ret']: agent.data_returns[i, :agent.top[i]]})
while True:
try:
x = agent.sess.run(agent.ops[i]['next_element'])
for j in x[4]:
data2[i][j] = data2[i][j] + 1 if j in data2[i] else 1
except tf.errors.OutOfRangeError:
break
assert data2 == data
# test learn
p1, p2 = xfp.get_random_policy()
agent.learn(10, [p1, p2], 100)
if __name__ == '__main__':
train()
# some_tests()