-
Notifications
You must be signed in to change notification settings - Fork 0
/
dqn.py
319 lines (224 loc) · 8.65 KB
/
dqn.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
from chainerrl.agents.dqn import DQN
from chainerrl import explorers
from chainerrl import links
from chainerrl import misc
from chainerrl import q_functions
from chainerrl import replay_buffer
from chainer import optimizers
import chainerrl
import logging
import sys
import os
import argparse
import chainer
import gym
# gym.undo_logger_setup() # NOQA
from gym import spaces
import gym.wrappers
import numpy as np
import marlo
from marlo import experiments
import time
games = ['MarLo-Obstacles-v0', 'MarLo-TrickyArena-v0', 'MarLo-Vertical-v0',
'MarLo-Attic-v0', 'MarLo-DefaultFlatWorld-v0', 'MarLo-DefaultWorld-v0',
'MarLo-Eating-v0', 'MarLo-CatchTheMob-v0', 'MarLo-CliffWalking-v0',
'MarLo-FindTheGoal-v0']
# For more details on DQN Parameters see https://chainerrl.readthedocs.io/en/latest/agents.html#agent-interfaces
# PARAMETER VARIABLES
GAME = games[9] # Game to run in
VIDEO_RES = 84 # Sets the resolution in pixels of the MARLO screen (at VIDEO_RES x VIDEO_RES)
DEBUG_ON = True # Limits output to console from this file
GPU = -1 # GPU to use (index: 0 to N). If no GPU available, set to -1.
# PARAMETER FUNCTIONS
# Sets the experimental profile
def set_experiment_profile():
# Number of total time steps (across all episodes) for training. When this number of played steps is reached, training is over
steps = 10 ** 6
# After how many episodes an evaluation is performed. Its results are dumped to "scores.txt"
eval_interval = 20 # 10 ** 4 # Commented values are default for large tasks.
# Number of times the game is played for each evaluation
eval_n_runs = 10 # 100
# Maximum duration for the episode during evaluation
max_eval_episode_len = 100
return steps, eval_n_runs, eval_interval, max_eval_episode_len
# Sets the discount factor for the network.
def set_discount_factor():
# Discount factor
gamma = 0.99
return gamma
# Sets the exploration dynamics.
def set_explorer(env):
# Possible parameters:
# Initial (and max) value of epsilon at the start of the experimentation.
start_epsilon = 1.0
# Minimum value of epsilon
end_epsilon = 0.1
# Constant epsilon
cons_epsilon = 0.001
# how many steps it takes for epsilon to decay
final_exploration_steps = 10 ** 5
# Options for exploration (more explorers at site-packages/chainerrl/explorers/)
# Option 1: Constant
constant_epsilon_explorer = explorers.ConstantEpsilonGreedy(
epsilon=cons_epsilon,
random_action_func=env.action_space.sample
)
# Option 2: Linear decay
decay_epsilon_explorer = explorers.LinearDecayEpsilonGreedy(
start_epsilon,
end_epsilon,
final_exploration_steps,
random_action_func=str(env.action_space.sample)
)
return constant_epsilon_explorer
# return decay_epsilon_explorer
def set_SDG_optimizer(q_function):
# Possible parameters:
alpha_learning_rate = 0.01
# Options for SDG (more optimizers available at site-packages/chainer/optimizers/ and https://docs.chainer.org/en/stable/reference/optimizers.html)
# Option 1: Vanilla Stochastic Gradient Descent
opt_sgd = optimizers.SGD(lr=alpha_learning_rate)
# Option 2: Adam (A Method for Stochastic Optimization) optimizer
opt_adam = optimizers.Adam()
opt_adam.setup(q_function)
return opt_sgd
#return opt_adam
# Sets the Replay Buffer to use for DQN.
def set_replay_buffer():
# Experience replay is disabled if the number of transitions in the replay buffer is lower than this value
replay_start_size = 1000
# After how many steps the batch of experiences are sampled again for SGD
update_interval = 1
# Capacity of the Reply Buffer for DQN
rbuf_capacity = 5 * 10 ** 5
return replay_start_size, update_interval, rbuf_capacity
# Sets the dynamics for updating the target network.
def set_network_target_update():
# Number of steps to update the target network
target_update_interval = 10 ** 2
# Method for updating the target weights: 'hard' or 'soft'
target_update_method = 'hard'
# Tau of soft target update
soft_update_tau = 1e-2
return target_update_interval, target_update_method, soft_update_tau
# ------------------------------------------------
# SUPPORT FUNCTIONS
# Sets the directory for output files.
def dirs(args):
out_dir = args.results_dir
save_dir = args.save_dir
if not os.path.exists(out_dir):
os.makedirs(out_dir)
out_dir_logs = out_dir + '/logging'
if not os.path.exists(out_dir_logs):
os.makedirs(out_dir_logs)
if save_dir and not os.path.exists(save_dir):
os.makedirs(save_dir)
return out_dir, save_dir
def phi(observation):
return observation.astype(np.float32)
# START!
# ARGUMENTS FOR THE DQN
parser = argparse.ArgumentParser(description='chainerrl dqn')
parser.add_argument('--n_hidden_channels', type=int, default=50, help='the number of hidden channels')
parser.add_argument('--n_hidden_layers', type=int, default=1, help='the number of hidden layers')
parser.add_argument('--results_dir', type=str, default="results", help='the results output dir')
parser.add_argument('--save_dir', type=str, default=None, help='the dir to save to or none')
parser.add_argument('--load_dir', type=str, default=None, help='the dir to save to or none.')
args = parser.parse_args()
out_dir, save_dir = dirs(args)
n_hidden_channels = args.n_hidden_channels
n_hidden_layers = args.n_hidden_layers
if DEBUG_ON:
print("n_hidden_channels " + str(n_hidden_channels) + " n_hidden_layers " + str(n_hidden_layers))
# GAME SELECTION AND CONNECTION TO THE CLIENT
# Ensure that you have a minecraft-client running with : marlo-server --port 10000
client_pool = [('127.0.0.1', 10020)]
if DEBUG_ON:
print("Game:", GAME)
join_tokens = marlo.make(GAME,
params=dict(
videoResolution=[VIDEO_RES, VIDEO_RES],
kill_clients_after_num_rounds=500
))
env = marlo.init(join_tokens[0])
# ------------------------------------------
obs = env.reset()
env.render(mode="rgb_array")
if DEBUG_ON:
print('initial observation:', obs)
action = env.action_space.sample()
obs, r, done, info = env.step(action)
if DEBUG_ON:
print('next observation:', obs)
print('reward:', r)
print('done:', done)
print('info:', info)
print('actions:', str(env.action_space))
print('sample action:', str(env.action_space.sample))
timestep_limit = env.spec.tags.get('wrapper_config.TimeLimit.max_episode_steps')
obs_space = env.observation_space
obs_size = obs_space.low.size
action_space = env.action_space
n_actions = action_space.n
q_func = q_functions.FCStateQFunctionWithDiscreteAction(
obs_size, n_actions,
n_hidden_channels=n_hidden_channels,
n_hidden_layers=n_hidden_layers
)
# Set up explorer
explorer = set_explorer(env)
# Set up replay buffer
replay_start_size, update_interval, rbuf_capacity = set_replay_buffer()
# Set up params for network target update
target_update_interval, target_update_method, soft_update_tau = set_network_target_update()
# Set up the Stochastic Gradient Descent Optimizer
optimizer = set_SDG_optimizer(q_func)
# Use GPU if any available
if GPU >= 0:
chainer.cuda.get_device(GPU).use()
q_func.to_gpu(GPU)
# DQN uses Experience Replay.
# Specify a replay buffer and its capacity.
rbuf = chainerrl.replay_buffer.ReplayBuffer(capacity=rbuf_capacity)
# Initialize the agent
agent = DQN(
q_func, optimizer, rbuf,
gpu=GPU,
gamma=set_discount_factor(),
explorer=explorer,
replay_start_size=replay_start_size,
target_update_interval=target_update_interval,
update_interval=update_interval,
phi=phi,
target_update_method=target_update_method,
soft_update_tau=soft_update_tau,
episodic_update_len=16
)
if args.load_dir:
if DEBUG_ON:
print("Loading model")
agent.load(args.load_dir)
# Sets the experiment profile
steps, eval_n_runs, eval_interval, max_eval_episode_len = set_experiment_profile()
# Trains an agent while regularly evaluating it.
experiments.train_agent_with_evaluation(
agent=agent,
env=env,
eval_env=env,
steps=steps,
eval_n_runs=eval_n_runs,
eval_interval=eval_interval,
outdir=out_dir,
max_episode_len=max_eval_episode_len, #timestep_limit
save_best_so_far_agent=False
)
if save_dir:
if DEBUG_ON:
print("Saving model")
agent.save(save_dir)
# Draw the computational graph and save it in the output directory.
chainerrl.misc.draw_computational_graph(
[q_func(np.zeros_like(obs_space.low, dtype=np.float32)[None])],
os.path.join(out_dir, 'model')
)