-
Notifications
You must be signed in to change notification settings - Fork 0
/
fedadmm_new_robot.py
436 lines (384 loc) · 20.7 KB
/
fedadmm_new_robot.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
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
import numpy as np
from scipy.linalg import solve_discrete_are
import cvxpy as cp
from ADMM_slack import *
from utils import *
import matplotlib.pyplot as plt
import warnings
# warnings.filterwarnings("ignore")
from datetime import datetime
import time
from scipy.stats import wishart
M = 1 # number of robots
n, m = 4, 2 # dimension of state (n) and input (m)
nrandom = 4 # number of random starts on ADMM (optimize over which random start gives the best L(K))
covarQR = 0.1 # covariance of generated Q and R (only matters if use_wishart = False)
niter = 10 # number of iterations within ADMM
Wdyn = 1 # magnitude of dynamics noise
Wctrl = 3 # magnitude of control noise
N = 5 # trajectory length
traj_range = np.arange(1, 16) # # of demonstrations (each of length N)
seed_range = np.arange(1, 11) # # of seeds to average over (the value actually doesn't matter, just the range length)
admm_QRreg = True # whether to add regularization on Q and R to Boyd's ADMM implementation
use_wishart = False # whether to initialize the Q's and R's using a wishart distribution
noise_on_output_loss = False
noise_on_input_data = True
average_over = 10
alpha = 0 # multiplier on r(K, Q, R)
alphaK = 0.05
beta = 0.01*100 # multiplier on l_D
name_clarifier = "ADMMslack_fixedrandom_newrobot"
# A = np.array([
# [.99997, .039, 0, -.322],
# [-.065, .99681, 7.74, 0],
# [.02, -.101, .99571, 0],
# [0, 0, 1, 1]
# ])
# B = np.array([
# [.0001, 0],
# [-.0018, -.0004],
# [-.0116, .00598],
# [0, 0]
# ])
# W = .1*np.array([
# [1.00092109, -0.02610491, 0.016055, 0.],
# [-0.02610491, 0.99785518, -0.10197781, 0.],
# [0.016055, -0.10197781, 0.010601, 0.],
# [0., 0., 0., 0.]
# ])
# A = np.random.randn(n, n)
# A = A/np.abs(np.linalg.eig(A)[0]).max()
# B = np.random.randn(n, m)
A = np.array([[0.13741826, 0.02876351, 0.50096039, 0.24711964],
[-0.08638155, -0.29891056, 0.16163153, 0.24719121],
[0.06460169, 0.24025215, -0.29127423, -0.28410344],
[0.18903618, 0.43957661, -0.43899608, -0.38508461]])
B = np.array([[1.12563572, -0.59218407],
[-0.01612495, -0.18979434],
[-1.38525459, -0.51277622],
[-1.43032195, 2.17007322]])
W = Wdyn * np.eye(n)
VQ = np.eye(n) / n # /n for normalizing to eye(n)
VR = np.eye(m) / m # /m
Q_trues = []
R_trues = []
P_trues = []
K_trues = []
ATA = lambda A: A.T @ A
Q_starter = np.array([[1.65775798, -0.12087396, 0.42784446, 0.7738255],
[-0.12087396, 0.15703623, 0.05796011, -0.09531927],
[0.42784446, 0.05796011, 0.1870196, 0.14953155],
[0.7738255, -0.09531927, 0.14953155, 0.40412161]])
R_starter = np.array([[1.25851049, 0.83728649],
[0.83728649, 3.77638641]])
for i in range(M):
if use_wishart:
Q = np.reshape(wishart.rvs(n * n, VQ), (n, n))
R = np.reshape(wishart.rvs(m * m, VR), (m, m))
else:
Q = Q_starter + covarQR * ATA(np.random.rand(n, n))
R = R_starter + covarQR * ATA(np.random.rand(m, m))
Q_trues.append(Q)
R_trues.append(R)
P_trues.append(solve_discrete_are(A, B, Q_trues[i], R_trues[i]))
K_trues.append(-np.linalg.solve(R_trues[i] + B.T @ P_trues[i] @ B, B.T @ P_trues[i] @ A))
print("Q_trues", Q_trues)
print("R_trues", R_trues)
print("K_trues", K_trues)
def simulate(K, robot, N=10, seed=None, add_noise=False, train=False):
if seed is None:
np.random.seed(np.random.randint(0, 1000))
x = np.random.multivariate_normal(np.zeros(n), W)
# if train:
# x = np.random.randint(-Wdyn*3, Wdyn*3, size=(n,))
# else:
# x = np.random.multivariate_normal(np.zeros(n), np.eye(n))
xs = []
us = []
cost = 0.0
for _ in range(N):
u = K @ x
# print('u', u)
if add_noise:
u += Wctrl * np.random.randn(m)
xs.append(x)
us.append(u)
cost += (x @ Q_trues[robot] @ x + u @ R_trues[robot] @ u) / N
# print('x', A@x + B@u)
x = A @ x + B @ u + np.random.multivariate_normal(np.zeros(n), W)
xs = np.array(xs)
us = np.array(us)
return cost, xs, us
def simulate_cost(K, robot, N=10, seed=None, add_noise=None, average_over=10, train=False):
if add_noise is None:
add_noise = noise_on_output_loss
tests = [simulate(K, robot, N=N, seed=seed, add_noise=add_noise, train=train) for _ in range(average_over)]
return np.nanmean([cost for cost, _, _ in tests])
N_test = 1000
# cost_true = np.nanmean([np.trace(P_trues[i]@W) for i in range(M)], axis=0)
# cost_noise = np.nanmean([np.trace(P_trues[0]@(W + Wctrl*Wctrl*[email protected])) for i in range(M)], axis=0)
cost_true = np.nanmean(
[simulate_cost(K_trues[i], i, N=N_test, add_noise=False, average_over=average_over) for i in range(M)], axis=0)
cost_noise = np.nanmean(
[simulate_cost(K_trues[i], i, N=N_test, add_noise=True, average_over=average_over) for i in range(M)], axis=0)
print("Cost true: {}, cost noise: {}".format(cost_true, cost_noise))
Q_avg = np.mean(Q_trues, axis=0)
R_avg = np.mean(R_trues, axis=0)
cost_LQ_true = np.nanmean([np.linalg.norm([ATA(1. / n * np.random.randn(n, n)) - Q_avg]) for _ in range(
N_test)], axis=0)
cost_LR_true = np.nanmean([np.linalg.norm([ATA(1. / n * np.random.randn(m, m)) - R_avg]) for _ in range(
N_test)], axis=0)
cost_fLQ_true = np.nanmean([np.linalg.norm(Q_trues[i] - Q_avg) for i in range(M)])
cost_fLR_true = np.nanmean([np.linalg.norm(R_trues[i] - R_avg) for i in range(M)])
print("Baseline error on LQ: {}, on LR: {}".format(cost_LQ_true, cost_LR_true))
print("Error for true federated learning on LQ: {}, on LR: {}".format(cost_fLQ_true, cost_fLR_true))
# For saving files
now = datetime.now()
timestamp = now.strftime("%Y%m%d_%H_%M_%S")
print(timestamp)
costs_admm = []
costs_pfedadmm = []
out_admm = {}
out_pfedadmm = {}
costs_lr_K = []
costs_admm_KQR = []
costs_pfedadmm_KQR = []
latexify(fig_width=6 * 2.5, fig_height=2.7 * 2.5)
def plot_losses(costs_admm, costs_pfedadmm, verbose=False, plot=False):
# global plotted_before, fig, axs
# fig.clf()
fig, axs = plt.subplots(2, 2)
costs_admm = np.array(costs_admm)
costs_pfedadmm = np.array(costs_pfedadmm)
np.save('data/' + timestamp + "_fedadmmv3_new_robot.npy", [A, B, K_trues, P_trues, Q_trues, R_trues,
costs_admm, costs_pfedadmm,
out_admm, out_pfedadmm,
costs_admm_KQR, costs_pfedadmm_KQR])
mean_admm = np.nanmean(costs_admm, axis=1)
std_admm = np.nanstd(costs_admm, axis=1)
mean_pfedadmm = np.nanmean(costs_pfedadmm, axis=1)
std_pfedadmm = np.nanstd(costs_pfedadmm, axis=1)
idx = np.arange(0, len(costs_lr_K))
idx_plot = np.arange(1, len(costs_lr_K) + 1)
mean_admm_KQR = {k: np.array([np.nanmean(costs_admm_KQR[i][k]) for i in idx]) for k in 'KQR'}
std_admm_KQR = {k: np.array([np.nanstd(costs_admm_KQR[i][k]) for i in idx]) for k in 'KQR'}
mean_pfedadmm_KQR = {k: np.array([np.nanmean(costs_pfedadmm_KQR[i][k]) for i in idx]) for k in 'KQR'}
std_pfedadmm_KQR = {k: np.array([np.nanstd(costs_pfedadmm_KQR[i][k]) for i in idx]) for k in 'KQR'}
if verbose:
print("Mean ADMM", mean_admm)
print("Mean pFedADMM", mean_pfedadmm)
axs[0, 0].axhline(cost_true, ls='-', c='k', label='optimal (without noise)')
# axs[0, 0].axhline(cost_noise, ls='--', c='k', label='expert (with noise)')
axs[0, 0].scatter(idx_plot, mean_admm, s=8, marker='*', c='green', label='ADMM')
axs[0, 0].fill_between(idx_plot, mean_admm - std_admm / 3, mean_admm + std_admm / 3, alpha=.3, color='green')
axs[0, 0].scatter(idx_plot, mean_pfedadmm, s=8, marker='*', c='purple', label='pFedADMM')
axs[0, 0].fill_between(idx_plot, mean_pfedadmm - std_pfedadmm / 3, mean_pfedadmm + std_pfedadmm / 3, alpha=.3,
color='purple')
axs[0, 0].semilogy()
axs[0, 0].set_ylabel('cost')
axs[0, 0].set_xlabel(r"# demonstrations $\tau_n$")
axs[0, 0].set_title("Cost vs. Trajectory Length")
axs[0, 0].legend()
# Plot K
axs[1, 0].scatter(idx_plot, mean_admm_KQR['K'], s=8, marker='o', c='green', label='ADMM')
axs[1, 0].fill_between(idx_plot, mean_admm_KQR['K'] - std_admm_KQR['K'] / 3,
mean_admm_KQR['K'] + std_admm_KQR['K'] / 3 / 3,
alpha=.3, color='green')
axs[1, 0].scatter(idx_plot, mean_pfedadmm_KQR['K'], s=8, marker='o', c='purple', label='pFedADMM')
axs[1, 0].fill_between(idx_plot, mean_pfedadmm_KQR['K'] - std_pfedadmm_KQR['K'] / 3,
mean_pfedadmm_KQR['K'] + std_pfedadmm_KQR['K'] / 3,
alpha=.3, color='purple')
axs[1, 0].grid(True)
axs[1, 0].set_xlabel(r"# demonstrations $\tau_n$")
axs[1, 0].set_ylabel(r'$||K - K_{true}||$')
axs[1, 0].set_title('K Loss, N=' + str(N) + ', M=' + str(M))
axs[1, 0].legend()
# Plot Q Loss
# axs[0, 1].axhline(cost_LQ_true, ls='-', c='k', label='Random Guessing')
axs[0, 1].axhline(cost_fLQ_true, ls='--', c='k', label='FedADMM with True Qavg')
axs[0, 1].scatter(idx_plot, mean_admm_KQR['Q'], s=8, marker='o', c='green', label='ADMM')
axs[0, 1].fill_between(idx_plot, mean_admm_KQR['Q'] - std_admm_KQR['Q'] / 3,
mean_admm_KQR['Q'] + std_admm_KQR['Q'] / 3 / 3,
alpha=.3, color='green')
axs[0, 1].scatter(idx_plot, mean_pfedadmm_KQR['Q'], s=8, marker='o', c='purple', label='pFedADMM')
axs[0, 1].fill_between(idx_plot, mean_pfedadmm_KQR['Q'] - std_pfedadmm_KQR['Q'] / 3,
mean_pfedadmm_KQR['Q'] + std_pfedadmm_KQR['Q'] / 3,
alpha=.3, color='purple')
axs[0, 1].grid(True)
axs[0, 1].set_xlabel(r"# demonstrations $\tau_n$")
axs[0, 1].set_ylabel(r'$||Q - Q_{true}||$')
axs[0, 1].set_title('Q Loss, N=' + str(N) + ', M=' + str(M))
axs[0, 1].legend()
# Plot R Loss
# axs[1, 1].axhline(cost_LR_true, ls='-', c='k', label='Random Guessing')
axs[1, 1].axhline(cost_fLR_true, ls='--', c='k', label='FedADMM with True Ravg')
axs[1, 1].scatter(idx_plot, mean_admm_KQR['R'], s=8, marker='o', c='green', label='ADMM')
axs[1, 1].fill_between(idx_plot, mean_admm_KQR['R'] - std_admm_KQR['R'] / 3,
mean_admm_KQR['R'] + std_admm_KQR['R'] / 3 / 3,
alpha=.3, color='green')
axs[1, 1].scatter(idx_plot, mean_pfedadmm_KQR['R'], s=8, marker='o', c='purple', label='pFedADMM')
axs[1, 1].fill_between(idx_plot, mean_pfedadmm_KQR['R'] - std_pfedadmm_KQR['R'] / 3,
mean_pfedadmm_KQR['R'] + std_pfedadmm_KQR['R'] / 3,
alpha=.3, color='purple')
axs[1, 1].grid(True)
axs[1, 1].set_xlabel(r"# demonstrations $\tau_n$")
axs[1, 1].set_ylabel(r'$||R - R_{true}||$')
axs[1, 1].set_title('R Loss, N=' + str(N) + ', M=' + str(M))
axs[1, 1].legend()
fig_name = "figures/" + timestamp + "_fedadmm_v2_random_M={}_Wctrl={}_Wdyn={}_nrandom={}_covarQR={}_niter=" \
"{}_nseed={}\n_N={}_Ntraj={}_admmQRreg={}_usewishartQR={}_" \
"noisyinput={}_noisyoutput={}" \
"_alpha={}_beta={}\n{}".format(
M, Wctrl, Wdyn, nrandom, covarQR, niter, len(seed_range), N, traj_range[-1], admm_QRreg, use_wishart,
noise_on_input_data, noise_on_output_loss, alpha, beta, name_clarifier
)
fig.suptitle(fig_name)
plt.tight_layout()
plt.savefig(fig_name.replace('\n', '') + ".png")
if not plot:
plt.close(fig)
for i in range(M):
for k in seed_range:
out_admm[(traj_range[0] - 1, k, i)] = (np.zeros((m, n)),
np.zeros((n, n)),
np.zeros((n, n)),
np.zeros((m, m))) # m, k, N --> (K, P, Q)
out_pfedadmm[(traj_range[0] - 1, k, i)] = (np.zeros((m, n)),
np.zeros((n, n)),
np.zeros((n, n)),
np.zeros((m, m))) # m, k, N --> (K, P, Q)
traj_prev = traj_range[0] - 1
xs_agg = {i: np.zeros((1, n)) for i in range(M)}
us_agg = {i: np.zeros((1, m)) for i in range(M)}
for traj in traj_range:
print("Traj # =", traj, end=" - ", flush=True)
# start = datetime.now()
start = time.time()
costs_admm += [[]]
costs_pfedadmm += [[]]
costs_lr_K += [{'K': []}]
costs_admm_KQR += [{'K': [], 'Q': [], 'R': []}]
costs_pfedadmm_KQR += [{'K': [], 'Q': [], 'R': []}]
for k in seed_range:
print(k, end=", (", flush=True)
prevfedQ = np.nanmean([out_admm[(traj_prev, k, i)][2] for i in range(M)], axis=0)
prevfedR = np.nanmean([out_admm[(traj_prev, k, i)][3] for i in range(M)], axis=0)
prevpfedQ = np.nanmean([out_pfedadmm[(traj_prev, k, i)][2] for i in range(M)], axis=0)
prevpfedR = np.nanmean([out_pfedadmm[(traj_prev, k, i)][3] for i in range(M)], axis=0)
if traj == 1:
prevfedP = np.zeros((n, n))
prevfedK = np.zeros((m, n))
prevpfedQ = Q_starter
prevpfedR = R_starter
else:
prevfedP = solve_discrete_are(A, B, prevfedQ, prevfedR)
prevfedK = -np.linalg.solve(prevfedR + B.T @ prevfedP @ B, B.T @ prevfedP @ A)
prevpfedP = solve_discrete_are(A, B, prevpfedQ, prevpfedR)
prevpfedK = -np.linalg.solve(prevpfedR + B.T @ prevpfedP @ B, B.T @ prevpfedP @ A)
# _, prevfedP, prevfedQ, prevfedR = out_pfedadmm[(traj_prev, k, i)]
for robot in range(M):
print(robot, end=", ", flush=True)
_, xs, us = simulate(K_trues[robot], robot, N=N, seed=k, add_noise=noise_on_input_data, train=True)
xs_agg[robot] = np.append(xs_agg[robot], xs, axis=0) # shape = N x n
us_agg[robot] = np.append(us_agg[robot], us, axis=0)
if traj == 1:
prevfedP = np.zeros((n, n))
prevfedK = np.zeros((m, n))
prevpfedP = np.zeros((n, n))
prevpfedK = np.zeros((m, n))
else:
prevpfedP = solve_discrete_are(A, B, prevpfedQ, prevpfedR)
prevpfedK = -np.linalg.solve(prevpfedR + B.T @ prevpfedP @ B, B.T @ prevpfedP @ A)
L = lambda K: cp.sum_squares(xs_agg[robot] @ K.T - us_agg[robot])
LPQR = lambda Q, R: beta * (cp.sum_squares(Q - prevfedQ) + cp.sum_squares(R - prevfedR))
r = lambda K: alphaK * cp.sum_squares(K)
rPQR = lambda Q, R: alpha * (cp.sum_squares(Q) + cp.sum_squares(R))
# 200*(cp.sum_squares(Q-Q_trues[robot]) + cp.sum_squares(R-R_trues[robot]))#
LK = lambda K: np.linalg.norm(K - K_trues[robot])
LQ = lambda Q: np.linalg.norm(Q - Q_trues[robot])
LR = lambda R: np.linalg.norm(R - R_trues[robot])
if admm_QRreg:
Kadmm, Padmm, Qadmm, Radmm = policy_fitting_with_a_kalman_constraint(L, r, A, B,
n_random=nrandom,
niter=niter,
P0=prevfedP,
Q0=prevfedQ,
R0=prevfedR,
rPQR=rPQR,
LPQR=LPQR)
else:
Kadmm, Padmm, Qadmm, Radmm = policy_fitting_with_a_kalman_constraint(L, r, A, B,
n_random=nrandom,
P0=prevfedP,
Q0=prevfedQ,
R0=prevfedR,
niter=niter)
# pFedADMM
LpPQR = lambda Q, R: beta * (cp.sum_squares(Q - prevpfedQ) + cp.sum_squares(R - prevpfedR))
Kpfedadmm, Ppfedadmm, Qpfedadmm, Rpfedadmm = policy_fitting_with_a_kalman_constraint(L, r,
A, B,
niter=niter,
n_random=nrandom,
P0=prevpfedP,
Q0=prevpfedQ,
R0=prevpfedR,
LPQR=LpPQR, rPQR=rPQR)
cost_admm = simulate_cost(Kadmm, robot, N=N_test, seed=0, average_over=average_over)
out_admm[(traj, k, robot)] = (Kadmm, Padmm, Qadmm, Radmm)
cost_pfedadmm = simulate_cost(Kpfedadmm, robot, N=N_test, seed=0, average_over=average_over)
out_pfedadmm[(traj, k, robot)] = (Kpfedadmm, Ppfedadmm, Qpfedadmm, Rpfedadmm)
if np.isnan(cost_admm) or cost_admm > 1e5 or cost_admm == np.inf:
cost_admm = np.nan
if np.isnan(cost_pfedadmm) or cost_pfedadmm > 1e5 or cost_pfedadmm == np.inf:
cost_pfedadmm = np.nan
costs_admm[-1].append(cost_admm)
costs_pfedadmm[-1].append(cost_pfedadmm)
costs_admm_KQR[-1]['K'].append(LK(Kadmm))
costs_admm_KQR[-1]['Q'].append(LQ(Qadmm))
costs_admm_KQR[-1]['R'].append(LR(Radmm))
costs_pfedadmm_KQR[-1]['K'].append(LK(Kpfedadmm))
costs_pfedadmm_KQR[-1]['Q'].append(LQ(Qpfedadmm))
costs_pfedadmm_KQR[-1]['R'].append(LR(Rpfedadmm))
loss_admm = np.linalg.norm(xs_agg[robot] @ Kadmm.T - us_agg[robot]) ** 2
loss_pfedadmm = np.linalg.norm(xs_agg[robot] @ Kpfedadmm.T - us_agg[robot]) ** 2
print('Ladmm = {}, Lpfedadmm = {}'.format(loss_admm,
loss_pfedadmm))
print(LQ(Q_starter), LR(R_starter))
print(LK(Kadmm), LQ(Qadmm), LR(Radmm))
print(LK(Kpfedadmm), LQ(Qpfedadmm), LR(Rpfedadmm))
avg_admmQ = np.zeros((n, n))
avg_admmR = np.zeros((m, m))
avg_pfedadmmQ = np.zeros((n, n))
avg_pfedadmmR = np.zeros((m, m))
for k in seed_range:
for robot in range(M):
avg_admmQ += out_admm[(traj, k, robot)][2] / len(seed_range) / M
avg_admmR += out_admm[(traj, k, robot)][3] / len(seed_range) / M
avg_pfedadmmQ += out_pfedadmm[(traj, k, robot)][2] / len(seed_range) / M
avg_pfedadmmR += out_pfedadmm[(traj, k, robot)][3] / len(seed_range) / M
true_admmQ_loss = np.linalg.norm(avg_admmQ - Q_avg)
true_admmR_loss = np.linalg.norm(avg_admmR - R_avg)
true_pfedadmmQ_loss = np.linalg.norm(avg_pfedadmmQ - Q_avg)
true_pfedadmmR_loss = np.linalg.norm(avg_pfedadmmR - R_avg)
end = time.time()
traj_prev = traj
print(
" %03d | %3.3f | %3.3f | "
"ADMM: %3.3f (%3.3f - %3.3f - %3.3f), LK = %3.3f, LQ = %3.3f (%3.3f), LR = %3.3f (%3.3f) | "
"pFedADMM: %3.3f (%3.3f - %3.3f - %3.3f), LK = %3.3f, LQ = %3.3f (%3.3f), LR = %3.3f (%3.3f) - "
"time elapsed = %3.3f " %
(N, cost_true, cost_noise,
np.nanmean(costs_admm[-1]), np.nanmin(costs_admm[-1]), np.nanstd(costs_admm[-1]), np.nanmax(costs_admm[-1]),
np.nanmean(costs_admm_KQR[-1]['K']),
np.nanmean(costs_admm_KQR[-1]['Q']), true_admmQ_loss,
np.nanmean(costs_admm_KQR[-1]['R']), true_admmR_loss,
np.nanmean(costs_pfedadmm[-1]), np.nanmin(costs_pfedadmm[-1]), np.nanstd(costs_pfedadmm[-1]),
np.nanmax(costs_pfedadmm[-1]),
np.nanmean(costs_pfedadmm_KQR[-1]['K']),
np.nanmean(costs_pfedadmm_KQR[-1]['Q']), true_pfedadmmQ_loss,
np.nanmean(costs_pfedadmm_KQR[-1]['R']), true_pfedadmmR_loss,
end - start
))
plot_losses(costs_admm, costs_pfedadmm, verbose=True)
plot_losses(costs_admm, costs_pfedadmm, verbose=True, plot=True)
plt.show()