-
Notifications
You must be signed in to change notification settings - Fork 1
/
sarsa_adagrad.py
304 lines (220 loc) · 11.7 KB
/
sarsa_adagrad.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
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
# Import
import numpy as np
import matplotlib.pyplot as plt
import numpy.matlib
import pandas as pd
from degree_freedom_queen import *
from degree_freedom_king1 import *
from degree_freedom_king2 import *
from generate_game import *
from Chess_env import *
import random
import os
import argparse
parser = argparse.ArgumentParser(description='Hyperparameter of SARSA algorithm')
parser.add_argument('--epsilon', type=float, help='Starting value of Epsilon for the Epsilon-Greedy policy', default=0.2)
parser.add_argument('--beta', type=float, help='The parameter sets how quickly the value of Epsilon is decaying',default=0.00005)
parser.add_argument('--gamma', type=float, help='The Discount Factor', default=0.85)
parser.add_argument('--eta', type=float, help='The Learning Rate', default=0.0035)
args = parser.parse_args()
## INITIALISE THE ENVIRONMENT
size_board = 4
env=Chess_Env(size_board)
def EpsilonGreedy_Policy(Qvalues, epsilon, allowed_a):
N_a=np.shape(Qvalues)[0]
a = list(range(0, N_a))
a = [i for idx, i in enumerate(a) if allowed_a[idx]]
Qvalues = [i for idx, i in enumerate(Qvalues) if allowed_a[idx]]
rand_value=np.random.uniform(0,1)
## epsilon is probability that we go random
rand_a=rand_value<epsilon
if rand_a==True:
a = random.choice(a)
else:
idx=np.argmax(Qvalues)
a = a[idx]
return a
def ComputeQvalues(W1, W2, bias_W1, bias_W2, X, hiddenactivfunction , outeractivfunction):
## Qvalues=np.matmul(W2, np.matmul(W1,X)) ## this is direct computation of hidden layer and then output layer, without applying any non linear activation function
## below is a better solution:
# Neural activation: input layer -> hidden layer
H1 = np.matmul(W1,X)+bias_W1 ## make sure that bias_W1 does not need to be transposed
# if hidden activ function is given:
if (hiddenactivfunction == 1):
H1 = np.round(1/(1+np.exp(-H1)), 5) ## sigmoid
elif(hiddenactivfunction == 2):
H1 = (H1>0).astype(int)*H1 ## RELU
Qvalues = np.matmul(W2,H1) + bias_W2
#if outer activ function is given
if (outeractivfunction == 1):
Qvalues = np.round(1/(1+np.exp(- Qvalues)), 5)
elif (outeractivfunction == 2):
Qvalues = (Qvalues>0).astype(int)*Qvalues
return Qvalues
# set seed
np.random.seed(2022)
# INITIALISE THE PARAMETERS OF YOUR NEURAL NETWORK
S,X,allowed_a=env.Initialise_game()
N_a=np.shape(allowed_a)[0] # TOTAL NUMBER OF POSSIBLE ACTIONS
N_in=np.shape(X)[0] ## INPUT SIZE
N_h=200 ## NUMBER OF HIDDEN NODES
## INITALISE YOUR NEURAL NETWORK... Here weights from input to hidden layer and from the hidden layer to output layer are initialized
W1 = np.random.randn(N_h, N_in) * np.sqrt(1 / (N_in))
W2 = np.random.randn(N_a, N_h) * np.sqrt(1 / (N_h))
bias_W1 = np.zeros((N_h,))
bias_W2 = np.zeros((N_a,))
# REWARD SCHEME DEFAULT
# Reward if episode is not ended: 0
# Reward if checkmate: 1
# Reward if draw: 0
N_episodes = 200 # THE NUMBER OF GAMES TO BE PLAYED 100000
hiddenactivfunction = 0
outeractivfunction = 1
# SAVING VARIABLES
R_save = np.zeros([N_episodes, 1])
N_moves_save = np.zeros([N_episodes, 1])
Delta_save = np.zeros([N_episodes, 1])
for n in range(N_episodes):
S,X,allowed_a=env.Initialise_game()
epsilon_f = args.epsilon / (1 + args.beta * n) ## DECAYING EPSILON
Done=0 ## SET DONE TO ZERO (BEGINNING OF THE EPISODE)
i = 1 ## COUNTER FOR NUMBER OF ACTIONS
if (n%10 ==0):
diagonal2 = np.zeros((N_h+1,N_a))
diagonal1 = np.zeros((N_in*N_h + N_h, ))
Qvalues = ComputeQvalues(W1, W2, bias_W1, bias_W2, X, hiddenactivfunction , outeractivfunction)
a=EpsilonGreedy_Policy(Qvalues,epsilon_f, allowed_a)
while (Done==0 and i<50): ## START THE EPISODE
Qvalues = ComputeQvalues(W1, W2, bias_W1, bias_W2, X, hiddenactivfunction , outeractivfunction)
S_next, X_next,allowed_a_next,R,Done=env.OneStep(a)
## THE EPISODE HAS ENDED, UPDATE...BE CAREFUL, THIS IS THE LAST STEP OF THE EPISODE
if Done==1:
R_save[n,]=np.copy(R)
N_moves_save[n,]=i
dEdQ=R-Qvalues[a]
Delta_save[n,] = 0.5*(dEdQ)*(dEdQ)
## update W2 and B2
dQdY = 1
if outeractivfunction == 1:
dYdQ = Qvalues[a]*(1-Qvalues[a])
elif outeractivfunction == 2:
dYdQ = (Qvalues[a]>0).astype(int)
H = np.matmul(W1,X) + bias_W1
if hiddenactivfunction == 1:
H = np.round(1/(1+np.exp(-H)), 5)
elif(hiddenactivfunction == 2):
H = (H>0).astype(int)*H
dYdW = H
gradient2 =dEdQ*dQdY*dYdW
grandientBias2 = dEdQ*dQdY
G = np.round(np.append(gradient2, grandientBias2), 4)
diagonal2[:,a] = diagonal2[:,a] + G**2
diagonal_to_use2 = (args.epsilon + diagonal2[:,a])**(-1/2)
W2[a,:]=W2[a,:]+args.eta*diagonal_to_use2[0:-1]*gradient2
bias_W2[a]=bias_W2[a]+args.eta*diagonal_to_use2[-1]*grandientBias2
## update W1 and B1 after W2 and B2 were updated
if hiddenactivfunction == 1:
dYdZ = (W2[a,:].reshape(1, 200) * H*(1-H).reshape(1, 200)).reshape(200,1)
elif(hiddenactivfunction == 2):
dYdZ = (W2[a,:].reshape(1, 200) * (H>0).astype(int)).reshape(200,1)
else:
dYdZ = W2[a,:].reshape(200, 1)
dZDW = X.reshape(1, 58)
gradient1 =dEdQ*dQdY*dYdZ*dZDW
grandientBias1 = dEdQ*dQdY*dYdZ.reshape(200,1)
G = np.ravel(np.concatenate((gradient1, grandientBias1), axis=1))
diagonal1 = diagonal1 + G**2
diagonal_to_use1 = ((args.epsilon+diagonal1).reshape(200, -1))**(-1/2)
W1[:,:]=W1[:,:]+ args.eta*diagonal_to_use1[:, 0:-1]*gradient1
bias_W1=bias_W1+ args.eta*diagonal_to_use1[:, -1]*grandientBias1.reshape(200,)
break
# IF THE EPISODE IS NOT OVER...
else:
Qvalues1=ComputeQvalues(W1, W2, bias_W1, bias_W2, X_next, hiddenactivfunction , outeractivfunction)
a1=EpsilonGreedy_Policy(Qvalues1,epsilon_f, allowed_a_next)
# Compute the delta
dEdQ=R+args.gamma*Qvalues1[a1] - Qvalues[a]
## update W2 and B2
dQdY = 1
if outeractivfunction == 1:
dYdQ = Qvalues[a]*(1-Qvalues[a])
elif outeractivfunction == 2:
dYdQ = (Qvalues[a]>0).astype(int)
H = np.matmul(W1,X) + bias_W1
if hiddenactivfunction == 1:
H = np.round(1/(1+np.exp(-H)), 5)
elif(hiddenactivfunction == 2):
H = (H>0).astype(int)*H
dYdW = H
gradient2 =dEdQ*dQdY*dYdW
grandientBias2 = dEdQ*dQdY
G = np.round(np.append(gradient2, grandientBias2), 4)
diagonal2[:,a] = diagonal2[:,a] + G**2
diagonal_to_use2 = (args.epsilon + diagonal2[:,a])**(-1/2)
W2[a,:]=W2[a,:]+args.eta*diagonal_to_use2[0:-1]*gradient2
bias_W2[a]=bias_W2[a]+args.eta*diagonal_to_use2[-1]*grandientBias2
## update W1 and B1 after W2 and B2 were updated
if hiddenactivfunction == 1:
dYdZ = (W2[a,:].reshape(1, 200) * H*(1-H).reshape(1, 200)).reshape(200,1)
elif(hiddenactivfunction == 2):
dYdZ = (W2[a,:].reshape(1, 200) * (H>0).astype(int)).reshape(200,1)
else:
dYdZ = W2[a,:].reshape(200, 1)
dZDW = X.reshape(1, 58)
gradient1 =dEdQ*dQdY*dYdZ*dZDW
grandientBias1 = dEdQ*dQdY*dYdZ.reshape(200,1)
G = np.ravel(np.concatenate((gradient1, grandientBias1), axis=1))
diagonal1 = diagonal1 + G**2
diagonal_to_use1 = ((args.epsilon+diagonal1).reshape(200, -1))**(-1/2)
W1[:,:]=W1[:,:]+ args.eta*diagonal_to_use1[:, 0:-1]*gradient1
bias_W1=bias_W1+ args.eta*diagonal_to_use1[:, -1]*grandientBias1.reshape(200,)
# NEXT STATE AND CO. BECOME ACTUAL STATE...
S=np.copy(S_next)
X=np.copy(X_next)
a = np.copy(a1)
i += 1 # UPDATE COUNTER FOR NUMBER OF ACTIONS
filename = os.path.splitext(__file__)[0]
save_path_dataframe_nmoves_ema="experiments/{0}/ep{1:.5f}_be{2:.5f}_ga{3:.5f}_et{4:.5f}/N_moves_SARSA_ema.csv".format(filename, args.epsilon, args.beta, args.gamma, args.eta)
save_path_dataframe_nmoves="experiments/{0}/ep{1:.5f}_be{2:.5f}_ga{3:.5f}_et{4:.5f}/N_moves_SARSA.csv".format(filename, args.epsilon,args.beta, args.gamma, args.eta)
# print(save_path_dataframe_nmoves)
save_path_plots_nmoves="experiments/{0}/ep{1:.5f}_be{2:.5f}_ga{3:.5f}_et{4:.5f}/nmoves.png".format(filename, args.epsilon, args.beta, args.gamma, args.eta)
save_path_dataframe_reward_ema="experiments/{0}/ep{1:.5f}_be{2:.5f}_ga{3:.5f}_et{4:.5f}/R_save_SARSA_ema.csv".format(filename, args.epsilon, args.beta, args.gamma, args.eta)
save_path_dataframe_reward="experiments/{0}/ep{1:.5f}_be{2:.5f}_ga{3:.5f}_et{4:.5f}/R_save_SARSA.csv".format(filename, args.epsilon, args.beta, args.gamma, args.eta)
save_path_plots_rsave="experiments/{0}/ep{1:.5f}_be{2:.5f}_ga{3:.5f}_et{4:.5f}/rsave.png".format(filename, args.epsilon, args.beta, args.gamma, args.eta)
save_path_dataframe_delta_ema="experiments/{0}/ep{1:.5f}_be{2:.5f}_ga{3:.5f}_et{4:.5f}/Delta_save_SARSA_ema.csv".format(filename, args.epsilon, args.beta, args.gamma, args.eta)
save_path_dataframe_delta="experiments/{0}/ep{1:.5f}_be{2:.5f}_ga{3:.5f}_et{4:.5f}/Delta_save_SARSA.csv".format(filename, args.epsilon, args.beta, args.gamma, args.eta)
save_path_plots_delta="experiments/{0}/ep{1:.5f}_be{2:.5f}_ga{3:.5f}_et{4:.5f}/delta.png".format(filename, args.epsilon, args.beta, args.gamma, args.eta)
# Plot the performance
N_moves_save_ema = pd.DataFrame(N_moves_save, columns = ['N_moves'])
N_moves_save_ema['N_moves'] = N_moves_save_ema['N_moves'].ewm(span=100, adjust=False).mean()
N_moves_save_ema.to_csv(save_path_dataframe_nmoves_ema)
N_moves_save = pd.DataFrame(N_moves_save, columns = ['N_moves'])
N_moves_save.to_csv(save_path_dataframe_nmoves)
plt.plot(N_moves_save_ema['N_moves'])
plt.xlabel('Episodes')
plt.ylabel('Number of Steps until "Done"')
plt.title('Average Number of Steps until "Done" per Episode')
# plt.show()
plt.savefig(save_path_plots_nmoves)
R_save_ema = pd.DataFrame(R_save, columns = ['R_save'])
R_save_ema['R_save'] = R_save_ema['R_save'].ewm(span=100, adjust=False).mean()
R_save_ema.to_csv(save_path_dataframe_reward_ema)
R_save = pd.DataFrame(R_save, columns = ['R_save'])
R_save.to_csv(save_path_dataframe_reward)
plt.plot(R_save_ema)
plt.xlabel('Episodes')
plt.ylabel('Reward')
plt.title('Average Rewards per Episode')
# plt.show()
plt.savefig(save_path_plots_rsave)
Delta_save_ema = pd.DataFrame(Delta_save, columns = ['Delta_save'])
Delta_save_ema['Delta_save'] = Delta_save_ema['Delta_save'].ewm(span=100, adjust=False).mean()
Delta_save_ema.to_csv(save_path_dataframe_delta)
Delta_save = pd.DataFrame(Delta_save, columns = ['Delta_save'])
Delta_save.to_csv(save_path_dataframe_delta)
plt.plot(Delta_save_ema)
plt.xlabel('Episodes')
plt.ylabel('Error')
plt.title('Average Loss')
# plt.show()
plt.savefig(save_path_plots_delta)