/
pokeenv.py
294 lines (242 loc) · 10 KB
/
pokeenv.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
import asyncio
import numpy as np
from gym.spaces import Box, Space
import logging
from gym.utils.env_checker import check_env
from poke_env.environment.abstract_battle import AbstractBattle
from poke_env.data.gen_data import GenData
from poke_env.player import (
Gen8EnvSinglePlayer,
RandomPlayer,
)
from pokeagent.agents.pokegym import PokeGen8Gym
from pokeagent.models.dqn import DQNAgent
from pokeagent.utils.reward import ShapedReward
def train_m1(env: PokeGen8Gym, agent: DQNAgent, episodes:int, sr:ShapedReward=None, device=None, save_dir=None):
"""
Training method 1: Sequential learning of reward function. Code is a bit jank but it gets the job done for now.
"""
META_STEPS = 5
shaped_reward = 0
shaped_reward_func = sr.generate_default_func()
for meta_step in range(META_STEPS):
for ep in range(episodes):
print('-=-=-=-=- NEW EP:', ep)
state, info = env.reset()
s, battle = state
steps = 0
average_loss = 0
while True:
# agent step and learn
action = agent.action(s) # [agent.action(state)]
new_state, reward, terminated, truncated, info = env.step(action)
new_s, new_battle = new_state[0], new_state[1]
done = terminated or truncated
shaped_reward = shaped_reward_func(battle, new_battle)
agent.cache(s, action, reward + shaped_reward, new_s, done)
q, loss = agent.optimize()
# logger.log_step(reward, loss, q)
# state = new_state
s = new_s
battle = new_battle
if done:
print('done!', done)
break
if loss is not None and loss > 0:
average_loss += loss
steps += 1
# log episode info
# logger.log_episode()
if ep > 0 and ep % 500 == 0:
evaluate(agent, 20)
if (steps > 0):
average_loss = average_loss / steps
else:
average_loss = -1
logging.info('average_loss', average_loss)
agent.save_all()
won, total_games = evaluate(agent, 20)
shaped_reward_func = sr.generate_reward_func(won / total_games)
agent = DQNAgent(embedding_size=env.input_size,
num_actions=env.action_space.n,
device=device,
evaluate=False,
lr=0.001,
save_dir=save_dir,
warmup=100,
name="iterate_{meta_step}")
sr.save()
env.close()
sr.save()
def train_m2(env: PokeGen8Gym, agent: DQNAgent, episodes:int, sr:ShapedReward=None, device=None, save_dir=None):
"""
Training method 2: Tree-based. Takes a while because I'm not using threading or async training...
"""
META_STEPS = 5
NUM_LEAVES = 5
shaped_reward = 0
shaped_reward_func = sr.generate_default_func()
for meta_step in range(META_STEPS):
MAX_REWARDS = []
for k in range(NUM_LEAVES):
for ep in range(episodes):
print('-=-=-=-=- NEW EP:', ep)
state, info = env.reset()
s, battle = state
steps = 0
average_loss = 0
while True:
# agent step and learn
action = agent.action(s) # [agent.action(state)]
new_state, reward, terminated, truncated, info = env.step(action)
new_s, new_battle = new_state[0], new_state[1]
done = terminated or truncated
shaped_reward = shaped_reward_func(battle, new_battle)
agent.cache(s, action, reward + shaped_reward, new_s, done)
q, loss = agent.optimize()
# logger.log_step(reward, loss, q)
# state = new_state
s = new_s
battle = new_battle
if done:
print('done!', done)
break
if loss is not None and loss > 0:
average_loss += loss
steps += 1
# log episode info
# logger.log_episode()
if ep > 0 and ep % 500 == 0:
evaluate(agent, 20)
if (steps > 0):
average_loss = average_loss / steps
else:
average_loss = -1
logging.info('average_loss', average_loss)
agent.save_all()
won, total_games = evaluate(agent, 20)
shaped_reward_func = sr.generate_reward_func(won / total_games)
agent = DQNAgent(embedding_size=env.input_size,
num_actions=env.action_space.n,
device=device,
evaluate=False,
lr=0.001,
save_dir=save_dir,
warmup=100,
name="iterate_{meta_step}_{k}")
MAX_REWARDS.append(won / total_games)
sr.save()
env.close()
sr.save()
def train_m3(env: PokeGen8Gym, agent: DQNAgent, episodes:int, sr:ShapedReward=None, device=None, save_dir=None):
"""
Training method 3
"""
shaped_reward = 0
shaped_reward_func = sr.generate_default_func() # sr.generate_reward_func([])
sr.save()
for ep in range(episodes):
print('-=-=-=-=- NEW EP:', ep)
state, info = env.reset()
s, battle = state
steps = 0
average_loss = 0
while True:
# agent step and learn
action = agent.action(s) # [agent.action(state)]
new_state, reward, terminated, truncated, info = env.step(action)
new_s, new_battle = new_state[0], new_state[1]
done = terminated or truncated
shaped_reward = shaped_reward_func(battle, new_battle)
agent.cache(s, action, reward + shaped_reward, new_s, done)
q, loss = agent.optimize()
# logger.log_step(reward, loss, q)
# state = new_state
s = new_s
battle = new_battle
if done:
print('done!', done)
break
if loss is not None and loss > 0:
average_loss += loss
steps += 1
# log episode info
# logger.log_episode()
if ep > 0 and ep % 500 == 0:
won, total_games = evaluate(agent, 20)
# shaped_reward_func = sr.generate_reward_func(won / total_games)
sr.save()
if (steps > 0):
average_loss = average_loss / steps
else:
average_loss = -1
logging.info('average_loss', average_loss)
env.close()
agent.save_all()
sr.save()
def evaluate(agent: DQNAgent, episodes:int):
eval_env = PokeGen8Gym(set_team=True, opponent="random") # change later
for ep in range(episodes):
state, info = eval_env.reset()
s, battle = state
while True:
# agent step and learn
action = agent.action(s) # [agent.action(state)]
new_state, reward, terminated, truncated, info = eval_env.step(action)
new_s, new_battle = new_state[0], new_state[1]
done = terminated or truncated
# state = new_state
s = new_s
if done:
logging.info(f'eval step {ep}/{episodes}')
print('done!', done)
break
logging.info(
f"DQN Evaluation: {eval_env.n_won_battles} victories out of {eval_env.n_finished_battles} episodes"
)
eval_env.close()
return eval_env.n_won_battles, eval_env.n_finished_battles
def evalw(agent: DQNAgent, eval_env, episodes:int):
for ep in range(episodes):
state, info = eval_env.reset()
s, battle = state
while True:
# agent step and learn
action = agent.action(s) # [agent.action(state)]
new_state, reward, terminated, truncated, info = eval_env.step(action)
new_s, new_battle = new_state[0], new_state[1]
done = terminated or truncated
# state = new_state
s = new_s
if done:
print('done!', done, reward)
break
logging.info(
f"DQN Evaluation: {eval_env.n_won_battles} victories out of {eval_env.n_finished_battles} episodes"
)
return eval_env.n_won_battles, eval_env.n_finished_battles
async def main():
# First test the environment to ensure the class is consistent
# with the OpenAI API
test_env = PokeGen8Gym(set_team=True, opponent="random")
# check_env(test_env)
test_env.close()
# Create one environment for training and one for evaluation
# opponent = RandomPlayer(battle_format="gen8randombattle")
train_env = PokeGen8Gym(set_team=True, opponent="random")
# opponent = RandomPlayer(battle_format="gen8randombattle")
# eval_env = SimpleRLPlayer(
# battle_format="gen8randombattle", opponent=opponent, start_challenging=True
# )
# Compute dimensions
n_action = train_env.action_space.n
input_shape = (1,) + train_env.observation_space.shape
n_steps = 10
done = False
while not done:
# Random action
action = train_env.action_space.sample()
(obs, battle), reward, done, info, what = train_env.step(action)
print(battle, obs, reward, done)
if __name__ == "__main__":
asyncio.get_event_loop().run_until_complete(main())