/
main_control.py
338 lines (281 loc) · 14 KB
/
main_control.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
"""
In tabular MDP setting, evaluates the learning of optimal policy using different guidance discount factors
On-policy means we run episodes,
in each episode we generate roll-outs/trajectories of current policy and run algorithm to improve the policy.
"""
import numpy as np
import matplotlib.pyplot as plt
import argparse
import timeit
import time
import ray
from copy import deepcopy
from utils.mdp_utils import MDP, SetMdpArgs
from utils.planing_utils import PolicyEvaluation, PolicyIteration, get_stationary_distrb, GetUniformPolicy
from utils.learning_utils import run_learning_method
from utils.common_utils import set_random_seed, create_result_dir, save_run_data, load_run_data, write_to_log, get_grid, \
start_ray, set_default_plot_params, save_fig, pretty_print_args
set_default_plot_params()
# -------------------------------------------------------------------------------------------
# Run mode
# -------------------------------------------------------------------------------------------
local_mode = False # True/False - run non-parallel to get error messages and debugging
# Option to load previous run results:
load_run = False # False/True If true just load results from dir, o.w run simulation
# result_dir_to_load = './saved/Multi_Exp/2020_05_30_07_40_26/Exp_5_LSTDQ_Discount_TV_dist'
result_dir_to_load = './saved/main_control/2020_06_28_11_59_58/'
# Plot options:
save_PDF = False # False/True - save figures as PDF file in the result folder
y_lim = None # [75, 85] | None
show_stars = True
# -------------------------------------------------------------------------------------------
# Set Parameters
# -------------------------------------------------------------------------------------------
args = argparse.Namespace()
# ----- Run Parameters ---------------------------------------------#
args.run_name = '' # 'Name of dir to save results in (if empty, name by time)'
args.seed = 1 # random seed
args.n_reps = 1000 # default 1000 # number of experiment repetitions
# how to create parameter grid:
# args.param_grid_def = {'type': 'gamma_guidance', 'spacing': 'linspace', 'start': 0.5, 'stop': 0.99, 'num': 50, 'decimals': 10}
args.param_grid_def = {'type': 'L2_factor', 'spacing': 'linspace', 'start': 1e-4, 'stop': 0.01, 'num': 50, 'decimals': 10}
args.config_grid_def = {'type': 'n_trajectories', 'spacing': 'list', 'list': [4, 8, 16, 32]} # grid of number of trajectories to generate per episode Default: [1, 2, 4, 8]
# args.config_grid_def = {'type': 'states_actions_TV_dist_from_uniform', 'spacing': 'list', 'list': [0, 0.3, 0.6, 0.9]} # note: above 0.9 is very slow
# args.config_grid_def = {'type': 'chain_mix_time', 'spacing': 'list', 'list': [5, 10, 20, 40, 80]} # must be above 1.0, generates states and actions from a Markov chain with some mixing time
# args.config_grid_def = {'type': 'n_episodes', 'spacing': 'list', 'list': [1, 2, 3, 4]}
# args.config_grid_def = {'type': 'None', 'spacing': 'list', 'list': [None]}
# ----- Problem Parameters ---------------------------------------------#
# MDP definition ( see data_utils.SetMdpArgs)
# args.mdp_def = {'type': 'RandomMDP', 'S': 10, 'A': 5, 'k': 2, 'reward_std': 0.1}
args.mdp_def = {'type': 'GridWorld', 'N0': 4, 'N1': 4, 'reward_std': 0.1, 'forward_prob_distrb': 'uniform', 'goal_reward': 1, 'R_low': -0.5, 'R_high': 0.5}
# args.mdp_def = {'type': 'GridWorld', 'N0': 4, 'N1': 4, 'reward_std': 0.1, 'forward_prob_distrb': {'alpha': 3, 'beta': 1}, 'goal_reward': 1}
args.depth = 10 # default: 10 # Length of trajectory
args.n_trajectories = 4 # default value
args.gammaEval = 0.99 # default: 0.99 # gammaEval
args.n_episodes = 5 # Number of episodes
args.train_sampling_def = {'type': 'Trajectories'} # n_traj trajectories of length args.depth
# args.train_sampling_def = {'type': 'Generative_uniform'} # n_samples = args.depth X n_traj
# args.train_sampling_def = {'type': 'Generative_Stationary'} # n_samples = args.depth X n_traj
# args.train_sampling_def = {'type': 'sample_all_s_a'} # n_samples per (s,a) = n_traj
# args.train_sampling_def = {'type': 'Generative', 'states_TV_dist_from_uniform': 0, 'actions_TV_dist_from_uniform': 0} # n_samples = args.depth X n_traj
# args.train_sampling_def = {'type': 'Generative', 'states_dist_from_uniform': 0.7, 'actions_dist_from_uniform': 0.7} # n_samples = args.depth X n_traj
# args.train_sampling_def = {'type': 'Generative', 'states_entropy': 1.5, 'actions_entropy': 1.0} # n_samples = args.depth X n_traj
# ----- Algorithm Parameters ---------------------------------------------#
args.initial_policy = 'uniform' # 'uniform' (default) | 'generated_random'
args.epsilon = 0.1 # for epsilon-greedy
args.default_gamma = None # default: None # The default guidance discount factor (if None use gammaEval)
args.default_l2_factor = 1e-4 # default: None # The default L2 factor (if using discount regularization) - note: it is necessary for LSTD
args.method = 'SARSA' # 'Policy-Evaluation Algorithm' # Options: 'Expected_SARSA' | 'Model_Based' | 'SARSA' | 'LSTDQ' | 'ELSTDQ' | 'ELSTDQ_nested' | 'LSTDQ_nested'
args.use_reward_scaling = False
# Hyper-parameters for iterative methods ('SARSA', 'Expected_SARSA' )
args.TD_Init_type = 'zero' # How to initialize V # Options: 'Vmax' | 'zero' | 'random_0_1' | 'random_0_Vmax' | '0.5_'Vmax' |
args.n_TD_iter = 5000 # Default: 500 for RandomMDP, 5000 for GridWorld # number of TD iterations
args.learning_rate_def = {'type': 'a/(b+i_iter)', 'a': 500, 'b': 1000, 'scale': True}
# -------------------------------------------------------------------------------------------
# A Ray remote function.
# Runs a single repetition of the experiment
@ray.remote
def run_rep(i_rep, alg_param_grid, config_grid_def, args_r):
set_random_seed(args_r.seed + i_rep)
config_grid_vals = get_grid(config_grid_def)
n_config_grid = len(config_grid_vals)
n_grid = len(alg_param_grid)
# runs a single repetition of the experiment
loss_rep = np.zeros((n_config_grid, n_grid))
gammaEval = args_r.gammaEval
config_type = config_grid_def['type']
# grid of number of trajectories to generate
for i_config, config_val in enumerate(config_grid_vals):
n_traj = args_r.n_trajectories
if config_type == 'n_trajectories':
n_traj = config_val
elif config_type == 'states_actions_TV_dist_from_uniform':
args_r.train_sampling_def = {'type': 'Generative', 'states_TV_dist_from_uniform': config_val, 'actions_TV_dist_from_uniform': config_val}
elif config_type == 'chain_mix_time':
args_r.train_sampling_def = {'type': 'chain_mix_time', 'mix_time': config_val}
elif config_type == 'n_episodes':
args_r.n_episodes = config_val
elif config_type == 'None':
pass
else:
raise AssertionError
# Generate MDP:
M = MDP(args_r)
# Optimal policy for the MDP:
pi_opt, V_opt, Q_opt = PolicyIteration(M, gammaEval)
# grid of regularization param
for i_grid, alg_param in enumerate(alg_param_grid):
gamma_guidance, l1_factor, l2_factor = get_regularization_params(args_r, alg_param,
args_r.param_grid_def['type'])
# run the learning episodes:
pi_t = run_learning_method(args_r, M, n_traj, gamma_guidance, l2_factor, l1_factor)
# Evaluate performance of learned policy:
V_t, _ = PolicyEvaluation(M, pi_t, gammaEval)
loss_rep[i_config, i_grid] = (np.abs(V_opt - V_t)).mean()
# end for grid
# end for i_config
return loss_rep
# end run_rep
# -------------------------------------------------------------------------------------------
def get_regularization_params(args_r, alg_param, reg_type):
gammaEval = args_r.gammaEval
if args_r.default_gamma is None:
gamma_guidance = gammaEval
else:
gamma_guidance = args_r.default_gamma
l2_factor = args.default_l2_factor
l1_factor = None
if reg_type in 'L2_factor':
l2_factor = alg_param
elif reg_type == 'L1_factor':
l1_factor = alg_param
elif reg_type == 'gamma_guidance':
gamma_guidance = alg_param
elif reg_type == 'NoReg':
pass
else:
raise AssertionError('Unrecognized regularization type: ' + str(reg_type))
return gamma_guidance, l1_factor, l2_factor
# -------------------------------------------------------------------------------------------
def run_simulations(args, save_result, local_mode, init_ray=True):
if init_ray:
start_ray(local_mode)
if save_result:
create_result_dir(args)
write_to_log('local_mode == {}'.format(local_mode), args)
start_time = timeit.default_timer()
set_random_seed(args.seed)
n_reps = args.n_reps
alg_param_grid = get_grid(args.param_grid_def)
n_grid = alg_param_grid.shape[0]
config_grid_vals = get_grid(args.config_grid_def)
n_config_grid = len(config_grid_vals)
planing_loss = np.zeros((n_reps, n_config_grid, n_grid))
info_dict = {}
# ----- Run simulation in parrnell process---------------------------------------------#
loss_rep_id_lst = []
for i_rep in range(n_reps):
# returns objects ids:
args_r = deepcopy(args)
planing_loss_rep_id = run_rep.remote(i_rep, alg_param_grid, args_r.config_grid_def, args_r)
loss_rep_id_lst.append(planing_loss_rep_id)
# end for i_rep
# ----- get the results --------------------------------------------#
for i_rep in range(n_reps):
loss_rep = ray.get(loss_rep_id_lst[i_rep])
if i_rep % max(n_reps // 100, 1) == 0:
time_str = time.strftime("%H hours, %M minutes and %S seconds",
time.gmtime(timeit.default_timer() - start_time))
write_to_log('Finished: {} out of {} reps, time: {}'.format(i_rep + 1, n_reps, time_str), args)
# end if
planing_loss[i_rep] = loss_rep
info_dict = {'planing_loss_avg': planing_loss.mean(axis=0), 'planing_loss_std': planing_loss.std(axis=0),
'alg_param_grid': alg_param_grid, 'n_reps_finished': i_rep + 1}
if save_result:
save_run_data(args, info_dict, verbose=0)
# end if
# end for i_rep
if save_result:
save_run_data(args, info_dict)
stop_time = timeit.default_timer()
write_to_log('Total runtime: ' +
time.strftime("%H hours, %M minutes and %S seconds", time.gmtime(stop_time - start_time)), args,
save_result)
return info_dict
# end run_simulations
# -------------------------------------------------------------------------------------------
def run_main_control(args, save_result=True, load_run_data_flag=False, result_dir_to_load='', save_PDF=False, plot=True,
local_mode=False, init_ray=True):
SetMdpArgs(args)
if load_run_data_flag:
args, info_dict = load_run_data(result_dir_to_load)
else:
info_dict = run_simulations(args, save_result, local_mode, init_ray=init_ray)
planing_loss_avg = info_dict['planing_loss_avg']
planing_loss_std = info_dict['planing_loss_std']
alg_param_grid = info_dict['alg_param_grid']
if 'n_reps_finished' in info_dict.keys():
n_reps_finished = info_dict['n_reps_finished']
else:
n_reps_finished = args.n_reps
# end if
# ----- Plot figures ---------------------------------------------#
if plot or save_PDF:
ax = plt.figure().gca()
if args.train_sampling_def['type'] in {'Generative', 'Generative_uniform', 'Generative_Stationary'}:
data_size_per_traj = args.depth * args.n_episodes
elif args.train_sampling_def['type'] == 'Trajectories':
data_size_per_traj = args.depth * args.n_episodes
elif args.train_sampling_def['type'] == 'sample_all_s_a':
data_size_per_traj = args.nS * args.nA * args.n_episodes
else:
raise AssertionError
# end if
xscale = 1.
legend_title = ''
if args.param_grid_def['type'] == 'L2_factor':
plt.xlabel(r'$L_2$ Regularization Factor [1e-2]')
xscale = 1e2
elif args.param_grid_def['type'] == 'L1_factor':
plt.xlabel(r'$L_1$ Regularization Factor ')
elif args.param_grid_def['type'] == 'gamma_guidance':
plt.xlabel(r'Guidance Discount Factor $\gamma$')
else:
raise AssertionError('Unrecognized args.grid_type')
# end if
ci_factor = 1.96 / np.sqrt(n_reps_finished) # 95% confidence interval factor
config_grid_vals = get_grid(args.config_grid_def)
for i_config, config_val in enumerate(config_grid_vals): # for of plots in the figure
if args.config_grid_def['type'] == 'n_trajectories':
if args.train_sampling_def['type'] in {'Generative_uniform', 'Generative', 'Generative_Stationary'}:
legend_title = 'Num. Samples'
label_str = '{} '.format(config_val * data_size_per_traj)
else:
legend_title = 'Num. Trajectories'
label_str = '{} '.format(config_val)
elif args.config_grid_def['type'] == 'states_actions_TV_dist_from_uniform':
legend_title = 'Total-Variation from\n uniform [normalized]'
label_str = '{} '.format(config_val)
elif args.config_grid_def['type'] == 'chain_mix_time':
legend_title = 'Mixing time'
label_str = '{} '.format(config_val)
elif args.config_grid_def['type'] == 'n_episodes':
legend_title = 'Num. Episodes'
label_str = '{} '.format(config_val)
else:
raise AssertionError
# end if
plt.errorbar(alg_param_grid * xscale, planing_loss_avg[i_config], yerr=planing_loss_std[i_config] * ci_factor,
marker='.', label=label_str)
if show_stars:
# Mark the lowest point:
i_best = np.argmin(planing_loss_avg[i_config])
plt.scatter(alg_param_grid[i_best] * xscale, planing_loss_avg[i_config][i_best], marker='*', s=400)
# end if
# for i_config
plt.grid(True)
plt.ylabel('Loss')
plt.legend(title=legend_title, loc='best', fontsize=12) # loc='upper right'
if y_lim:
plt.ylim(y_lim)
# plt.xlim([0.5,1])
# ax.set_yticks(np.arange(0., 9., step=1.))
# plt.figure(figsize=(5.8, 3.0)) # set up figure size
if save_PDF:
save_fig(args.run_name)
else:
# plt.title('Loss +- 95% CI \n ' + str(args.args))
plt.title(args.mdp_def['type'] + ' ' + args.run_name + ' \n ' + args.result_dir, fontsize=6)
# end if save_PDF
# end if
pretty_print_args(args)
if plot:
plt.show()
print('done')
info_dict['result_dir'] = args.result_dir
return info_dict
# end run_main
# -------------------------------------------------------------------------------------------
if __name__ == "__main__":
info_dict = run_main_control(args, save_result=True, load_run_data_flag=load_run,
result_dir_to_load=result_dir_to_load, save_PDF=save_PDF, local_mode=local_mode)