-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmain.py
150 lines (143 loc) · 6.2 KB
/
main.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
#!/usr/bin/env python
# -*- coding: utf-8 -*-
'''
@Author: qinrj
@Description:
@Date: 5/19/18 7:29 PM
@Contact: qinrj@lamda.nju.edu.cn or 2428921608@qq.com
'''
from __future__ import print_function
from __future__ import division
import ConfigParser
from environment import GridRoom
from time import *
import numpy as np
from copy import *
from agent import *
import cPickle
config = ConfigParser.ConfigParser()
with open('config.cfg', 'rw') as cfgfile:
config.readfp(cfgfile)
_width = int(config.get('environ', 'Width'))
_height = int(config.get('environ', 'Height'))
_num_players = int(config.get('environ', 'Players'))
_ammo = int(config.get('environ', 'Ammo'))
_max_step = int(config.get('environ', 'Max_Step'))
_gamma = float(config.get('environ', 'Gamma'))
def train():
world = GridRoom()
players = [RMAgent(i) for i in range(0, _num_players)]
begin = time()
# sampled_exp = []
for _iteration in range(100000):
sampled_exp = []
iter_time = time()
# for _pid in range(1, _num_players):
# players[_pid].u_s = copy(players[0].u_s)
# players[_pid].u_sa = copy(players[0].u_sa)
# players[_pid].average_strategy = copy(players[0].average_strategy)
for _i in range(1000):
for _pid in range(_num_players):
players[_pid].set_ammo(_num_players - 1)
players[_pid].exp_buffer = []
ob = world.cur_state()
done = False
prev_j_ac = [9] * _num_players
while not done:
joint_action = [9] * _num_players
for _pid in world.alive_players:
_ac = players[_pid].action([ob, prev_j_ac], players[_pid].valid_action())
if _ac == 0:
players[_pid].set_ammo(players[_pid].ammo - 1)
joint_action[_pid] = _ac
done, r, n_state = world.step(joint_action)
for _pid in world.alive_players:
players[_pid].exp_buffer.append([str(_pid) + ''.join(map(str, ob)) +
str(players[_pid].ammo) + ''.join(map(str, prev_j_ac)), joint_action[_pid], r[_pid]])
for _pid in world.dead_in_this_step:
players[_pid].exp_buffer.append([str(_pid) + ''.join(map(str, ob)) +
str(players[_pid].ammo) + ''.join(map(str, prev_j_ac)), joint_action[_pid], r[_pid]])
# for _pid in world.dead_players:
# if r[_pid]
prev_j_ac = copy(joint_action)
ob = n_state
for _pid in range(_num_players): # TODO : replace with player 1
v = 0
for _k in range(len(players[_pid].exp_buffer) - 1, -1, -1):
players[_pid].exp_buffer[_k][2] += _gamma * v
v = players[_pid].exp_buffer[_k][2]
# if v < 0:
# print(players[_pid].exp_buffer)
# print(players[_pid].exp_buffer)
sampled_exp.extend(players[_pid].exp_buffer) # all agents share it if sp is used
world.reset()
print('Episode time: %.2f' % (time() - iter_time))
# end sample
update_time = time()
for _pid in range(_num_players):
players[_pid].update_policy(sampled_exp)
print('Update time: %.2f' %(time() - update_time))
if (_iteration + 1) % 100 == 0:
print('This is %d step, %.3f' % (_iteration, _iteration/100000.0))
if (_iteration + 1) % 10000 == 0:
for _pid in range(_num_players):
with open('v{}_{}.pkl'.format(_pid, _iteration / 10000), 'wb') as f:
cPickle.dump(players[0].u_s, f, 2)
with open('q{}_{}.pkl'.format(_pid, _iteration / 10000), 'wb') as f:
cPickle.dump(players[0].u_sa, f, 2)
with open('pi{}_{}.pkl'.format(_pid, _iteration / 10000), 'wb') as f:
cPickle.dump(players[0].average_strategy, f, 2)
print('Time eplapsed: %.2f' % (time() - begin))
def test():
world = GridRoom()
players = [RMAgent(0)]
for i in range(1, _num_players):
players.append(RandomAgent(i))
players[0].test = True
begin = time()
with open('v0.pkl', 'rb') as f:
players[0].u_s = cPickle.load(f)
with open('q0.pkl', 'rb') as f:
players[0].u_sa = cPickle.load(f)
with open('pi0.pkl', 'rb') as f:
players[0].average_strategy = cPickle.load(f)
total_r = np.zeros(_num_players)
print('Time for load model: ', time()-begin, players[0].u_s.__len__(), players[0].u_sa.__len__(), players[0].average_strategy.__len__())
begin = time()
# sampled_exp = []
for _iteration in range(100000):
iter_time = time()
for _i in range(1):
for _pid in range(_num_players):
players[_pid].set_ammo(_num_players - 1)
ob = world.cur_state()
done = False
prev_j_ac = [9] * _num_players
while not done:
joint_action = [9] * _num_players
for _pid in world.alive_players:
_ac = players[_pid].action([ob, prev_j_ac], players[_pid].valid_action())
if _ac == 0:
players[_pid].set_ammo(players[_pid].ammo - 1)
joint_action[_pid] = _ac
done, r, n_state = world.step(joint_action)
prev_j_ac = copy(joint_action)
ob = n_state
total_r = np.add(total_r, world.players_total_reward)
world.reset()
# print('Episode time: %.2f' % (time() - begin))
print('Time eplapsed: %.2f min' % ((time() - begin)/60.0))
print(total_r)
print('unseen, seen', players[0].unseen, players[0].seen)
# for _key, _item in players[0].u_s.iteritems():
# print(_key, _item)
# print('------------')
# for _key, _item in players[0].u_sa.iteritems():
# print(_key, _item)
# print('------------')
# for _k, _i in players[0].average_strategy.iteritems():
# if _i.count(0) < _i.__len__()-1:
# print(_k, _i)
if __name__ == '__main__':
train()
# test()