-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathRH_main_fp.py
172 lines (140 loc) · 9.21 KB
/
RH_main_fp.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
163
164
165
166
167
168
169
import subprocess
import numpy as np
import args_parser
from env.fast_marl import FastMARLEnv
from utils import get_softmax_action_probs_from_Qs, get_action_probs_from_Qs, get_curr_mf, find_best_response, \
eval_curr_reward,find_soft_response, get_softmax_new_action_probs_from_Qs, get_new_action_probs_from_Qs
if __name__ == '__main__':
config = args_parser.parse_config(mf_method="RH")
#Change directory so it differs from original method
# Unpack config
# game = config['game']
# fp_iterations = config['fp_iterations']
# variant = config['variant']
tau = config['tau']
# temperature = config['temperature']
# method = config['method']
env: FastMARLEnv = config['game'](**config)
mu_0 = env.mu_0
total_horizon = env.time_steps
action_probs_final = np.zeros((env.time_steps, env.observation_space.n, env.action_space.n))
mu_final = np.zeros((env.time_steps+1, env.observation_space.n))
mu_final[0] = mu_0
with open(config['exp_dir'] + f"stdout", "w", buffering=1) as fo:
for i in range(total_horizon-tau+1):
#Run MFG process, with time_steps = tau, and the initial condition for the mean field by calling main
# p = subprocess.Popen(['python',
# './main_fp.py',
# f'--game={game}',
# f'--fp_iterations={fp_iterations}',
# f'--method={method}',
# f'--variant={variant}',
# f'--tau={tau}',
# f'--temperature={temperature}',
# f'--time_steps={tau}',
# #f'--mu_0={mu_0}'
# ])
# Copy of main
#Initial
env.mu_0 = mu_final[i]
env.time_steps = tau
Q_0 = np.zeros((env.time_steps, env.observation_space.n, env.action_space.n))
action_probs = get_action_probs_from_Qs(Q_0)
#For FP
sum_action_probs = np.zeros_like(action_probs)
mus_avg = get_curr_mf(env, action_probs)
beta = 0.95
""" Compute the MFG fixed point for all high degree agents """
for iteration in range(config['fp_iterations']):
if config['method'] == 'pFP':
#FP method where the policy gets averaged.
sum_action_probs = sum_action_probs * beta + (1-beta)*action_probs
action_probs = sum_action_probs/sum_action_probs.sum(-1)[...,None]
action_probs[np.isnan(action_probs)] = 1 / env.action_space.n
mus = get_curr_mf(env, action_probs)
"""Evaluation"""
# IF FP: we have to compare the average policy against the best response to the meanfield induced by the average policy
# This is only the average policy, that leads to the average mean field, when the dynamics do not depend on the mean field
if config['method'] == "FP":
sum_action_probs = sum_action_probs + action_probs * mus[:-1][..., None]
action_probs_avg = sum_action_probs/sum_action_probs.sum(-1)[...,None]
action_probs_avg[np.isnan(action_probs_avg)] = 1 / env.action_space.n
action_probs_compare = action_probs_avg.copy()
mu_compare = get_curr_mf(env, action_probs_compare)
elif config['method'] == "expFPv2":
if iteration ==0:
sum_action_probs = action_probs * mus[:-1][..., None]
else:
sum_action_probs = sum_action_probs * beta + (1-beta)*action_probs * mus[:-1][..., None]
action_probs_avg = sum_action_probs / sum_action_probs.sum(-1)[..., None]
action_probs_avg[np.isnan(action_probs_avg)] = 1 / env.action_space.n
action_probs_compare = action_probs_avg.copy()
mu_compare = get_curr_mf(env, action_probs_compare)
else:
action_probs_compare = action_probs.copy()
mu_compare = mus.copy()
""" Evaluate current policy """
V_pi, Q_pi = eval_curr_reward(env, action_probs_compare, mu_compare)
""" Evaluate current best response against current average policy """
Q_br = find_best_response(env, mu_compare)
v_1 = np.vdot(env.mu_0, Q_br.max(axis=-1)[0])
v_curr_1 = np.vdot(env.mu_0, V_pi)
""" Exploitability """
print(f"{config['exp_dir']} game {i} iteration {iteration}: expl: {v_1 - v_curr_1}, ... br achieves {v_1} vs. {v_curr_1}")
fo.write(f"{config['exp_dir']} game {i} iteration {iteration}: expl: {v_1 - v_curr_1}, ... br achieves {v_1} vs. {v_curr_1}")
fo.write('\n')
""" Compare Policies """
"""Boltzmann L1-Distance """
BE_action_probs = get_softmax_action_probs_from_Qs(Q_br, temperature=config['temperature'])
print(f"{config['exp_dir']} game {i} iteration {iteration}: BE_l1_distance: {np.abs(BE_action_probs - action_probs_compare).sum(-1).sum(-1).max()}")
fo.write(f"{config['exp_dir']} game {i} iteration {iteration}: BE_l1_distance: {np.abs(BE_action_probs - action_probs_compare).sum(-1).sum(-1).max()}")
fo.write('\n')
"""QRE L1-Distance"""
QRE_action_probs = get_softmax_action_probs_from_Qs(Q_pi, temperature=config['temperature'])
print(f"{config['exp_dir']} game {i} iteration {iteration}: QRE_l1_distance: {np.abs(QRE_action_probs - action_probs_compare).sum(-1).sum(-1).max()}")
fo.write(f"{config['exp_dir']} game {i} iteration {iteration}: QRE_l1_distance: {np.abs(QRE_action_probs - action_probs_compare).sum(-1).sum(-1).max()}")
fo.write('\n')
"""Relative Entropy L1-Distance"""
Q_sr = find_soft_response(env, mu_compare, temperature=config['temperature'])
RE_action_probs = get_softmax_action_probs_from_Qs(Q_sr, temperature=config['temperature'])
print(f"{config['exp_dir']} game {i} iteration {iteration}: RE_l1_distance: {np.abs(RE_action_probs - action_probs_compare).sum(-1).sum(-1).max()}")
fo.write(f"{config['exp_dir']} game {i} iteration {iteration}: RE_l1_distance: {np.abs(RE_action_probs - action_probs_compare).sum(-1).sum(-1).max()}")
fo.write("\n")
### Average mean_field for FP methods
if config['method']=='FP':
mus_avg = (iteration * mus_avg + mus) / (iteration + 1)
elif (config['method']=='expFPv1')|(config['method']=='expFPv2'):
mus_avg = beta * mus_avg + (1 - beta) * mus
#Nash Equilibrium
if config['variant'] == 'NE':
#If FP method compute best response wtr mus_avg
if (config['method'] == 'FP') | (config['method']=='expFPv1')|(config['method']=='expFPv2'):
Q_br = find_best_response(env, mus_avg)
action_probs = get_action_probs_from_Qs(Q_br)
elif config['variant'] == 'QRE':
#If FP method compute Q_pi wtr mus_avg
if (config['method'] == 'FP') | (config['method']=='expFPv1')|(config['method']=='expFPv2'):
V_pi, Q_pi = eval_curr_reward(env, action_probs, mus_avg)
action_probs = get_softmax_action_probs_from_Qs(Q_pi, temperature=config['temperature'])
elif config['variant'] == 'BE':
#If FP method compute best response wtr mus_avg
if (config['method'] == 'FP') |(config['method']=='expFPv1')|(config['method']=='expFPv2'):
Q_br = find_best_response(env, mus_avg)
action_probs = get_softmax_action_probs_from_Qs(Q_br, temperature=config['temperature'])
elif config['variant'] == "RE":
#If FP method compute soft response wtr mus_avg
if (config['method'] == 'FP') |(config['method']=='expFPv1')|(config['method']=='expFPv2'):
Q_sr = find_soft_response(env,mus_avg,temperature=config['temperature'])
elif (config['method']=="FPI")|(config['method']=='pFP'):
Q_sr = find_soft_response(env, mus, temperature=config['temperature'])
action_probs = get_softmax_action_probs_from_Qs(Q_sr, temperature=config['temperature'])
else:
raise NotImplementedError
action_probs_final[i] = action_probs_compare[0]
mu_final[i+1] = mu_compare[1]
#The last tau actions are taken from one mfg
action_probs_final[i:] = action_probs_compare[0:]
mu_final[i + 1:] = mu_compare[1:]
np.save(config['exp_dir'] + f"action_probs.npy", action_probs_final)
np.save(config['exp_dir'] + f"best_response.npy", Q_br)
np.save(config['exp_dir'] + f"mean_field.npy", mu_final)