forked from tobspr-games/shapez.io
-
Notifications
You must be signed in to change notification settings - Fork 0
/
model.py
548 lines (444 loc) · 19.3 KB
/
model.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
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
# -*- coding: utf-8 -*-
"""
Created on Tue Aug 13, 2024 at 09:55:41
@author: Cain Bruhn-Tanzer, Rhys Tyne
"""
import os
import sys
# Set environment flags before running imports.
os.environ["KERAS_BACKEND"] = "tensorflow"
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2' # Supress INFO Logs
# pylint: disable=C0413
from datetime import datetime
import json
import random
import logging
import numpy as np
import tensorflow as tf
import keras
# Configure logging
log = logging.getLogger(__name__)
logging.basicConfig(
level=logging.DEBUG,
format='%(asctime)s:%(levelname)s: %(message)s',
datefmt="%H:%M:%S",
handlers=[logging.StreamHandler()]
)
GOALS = [['CuCuCuCu', 30]] # currently only level 1
MODEL_STORAGE_DIRECTORY = "./src_ai/models"
class Model:
""" An abstract model class for AI development. """
def __init__(self):
self.alive = True
self.name = "AbstractModel"
self.version = "0.0.0"
self.model = {}
def save(self, obj):
""" Saves the current model as a JSON schema file. """
os.makedirs(MODEL_STORAGE_DIRECTORY, exist_ok=True)
folder = MODEL_STORAGE_DIRECTORY
name = self.name
version = self.version
dtime = datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
path = f"{folder}/{name}_{version}_{dtime}.json"
with open(path, 'w', encoding='utf-8') as json_file:
json.dump(obj, json_file, indent=4)
def load(self, file):
""" Loads a model saved as a JSON schema file. """
with open(file, 'r', encoding='utf-8') as json_file:
self.model = json.load(json_file)
def is_alive(self):
""" Returns whether model is running. """
return self.alive
def get_name(self):
""" Returns the name of the current model. """
return self.name
def train(self, game): # pylint: disable=W0613
""" Advances the model multiple steps to complete a training cycle. """
return None
def validate(self):
""" Checks a given solution and assigns points to it. """
return None
def _step(self, game): # pylint: disable=W0613
""" Advances the model one step. """
return True
def query(self, scenario): # pylint: disable=W0613
""" Queries the Model for a solution to a specific scenario. """
return None
class Overseer(Model):
""" Model for optimising higher level logistics networks.
Requirements:
- Must be able to draw and optimise networks between existing nodes
where nodes is a sub-factory layout.
- Must maintain a collection of possible node choices generated by
architects.
"""
def __init__(self, seed=1234):
super().__init__()
self.name = "Overseer"
self.version = "0.1.0"
self.seed = seed
self.nodes = []
def query(self, scenario):
""" Returns a single action given a scenario.
- Build a response from the overseer
- Add in a few queries of Architect if necessary.
"""
# Trigger something in the model based on the gameState
temp_response = []
# Place a strip of belts
temp_response.extend([
{"type": "Belt", "x": 2, "y": y, "rotation": 270}
for y in range(-2, 2)]
)
# Place a strip of readers
temp_response.extend([
{"type": "Reader", "x": 3, "y": y, "rotation": 270}
for y in range(-2, 2)]
)
# Place a strip of miners
temp_response.extend([
{"type": "Miner", "x": 4, "y": y, "rotation": 270}
for y in range(-2, 2)]
)
temp_response.extend([
{"type": "Miner", "x": 5, "y": y, "rotation": 270}
for y in range(-2, 1)]
)
temp_response.extend([
{"type": "Miner", "x": 6, "y": y, "rotation": 270}
for y in range(-2, 0)]
)
temp_response.extend([
{"type": "Miner", "x": 7, "y": -2, "rotation": 270}
])
return temp_response
class Architect(Model):
""" Model for designing sub-unit factory layouts. """
def __init__(self, seed=42):
super().__init__()
self.name = "Architect"
self.version = "0.4.0"
self.state_machine = "ONLINE"
# The current state values
self.region = None
self.action_space = None # Dict({ "f'{x}|{y}'": ["ABCD"], etc })
self.num_actions = 0
self.queued_action = None
# Model Training Factors
self.seed = seed
self.model = {}
self.target = {}
self.optimiser = keras.optimizers.Adam(learning_rate=0.0001)
self.episodes = 0
self.max_episodes = 10 # TODO was 10
self.max_frames = 20 # TODO was 10,000
self.running_reward = 0
self.episode_reward = 0
self.frames = 0
# Chosen Hyperparameters
self.gamma = 0.99
self.epsilon = 1.0
self.epsilon_min = 0.1
self.epsilon_max = 1.0
self.epsilon_interval = self.epsilon_max - self.epsilon_min
self.batch_size = 32
# Training Values
self.epsilon_random_frames = 50000 # Random Action Frames
self.epsilon_greedy_frames = 1000000.0 # Exploration Frames
self.max_memory_length = 100 # Maximum replay length
# TODO: Deepmind Suggests 100,000 memory max, significant memory usage
self.update_after_actions = 4 # Train Model every X Actions
self.update_target_network = 10000 # Network Update Target
self.loss_function = keras.losses.Huber() # huber loss for stability
# Experience replay buffers
self.action_history = []
self.state_history = []
self.state_next_history = []
self.goal_history = []
self.rewards_history = []
self.episode_reward_history = []
def get_state_machine(self):
""" Returns the current state in the state machine. """
return self.state_machine
def create_q_model(self, actions):
""" Creates a Deep Q Style Model as seen in the deepmind paper. """
# See https://keras.io/examples/rl/deep_q_network_breakout/
return keras.Sequential(
[
keras.layers.Input(shape=(4, 84, 84)),
keras.layers.Lambda(
lambda tensor: keras.ops.transpose(tensor, [0, 2, 3, 1]),
output_shape=(84, 84, 4),
),
# Convolutions on the frames on the screen
keras.layers.Conv2D(32, 8, strides=4, activation="relu",),
keras.layers.Conv2D(64, 4, strides=2, activation="relu"),
keras.layers.Conv2D(64, 3, strides=1, activation="relu"),
keras.layers.Flatten(),
keras.layers.Dense(512, activation="relu"),
keras.layers.Dense(actions, activation="linear"),
]
)
def _get_training_status(self):
""" Utility: Gets the current training status. """
if self.get_state_machine == "ONLINE":
return "Not Training."
e, e_max = (self.episodes, self.max_episodes)
f, f_max = (self.frames, self.max_frames)
a = str(self.queued_action).ljust(20)
return f" -> Training: Ep {e}|{e_max}, Fr {f}|{f_max}, Ac {a}"
def train(self, game):
""" Begins the model training state machine.
NOTE: Frankly I hate that this is necessary, but keeping the two
systems in lockstep is painful.
1. Train: Sets the GameState as a region and action space.
2. Episode: Starts an episode that triggers a reset on the client.
3. Pre-Frame: Queues an action for the client to apply to the game.
-> Client Runs Action for X Steps to generate new state for frame.
4. Post-Frame: New state is validated.
"""
# 1. Train
if self.get_state_machine() == "ONLINE":
print(" -> Training Request")
self.region = game.get_region(-18, -18, 36, 36)
self.action_space = game.get_action_space(self.region)
self.num_actions = len(self.action_space)
print(" -> Setting Key Values")
self.seed = game.get_seed()
self.episode_reward = 0
print(" -> Creating Deep Q Model & Target Model")
# TODO Define Model outside (and before) of training loop
self.model = self.create_q_model(self.num_actions)
self.target = self.create_q_model(self.num_actions)
self.state_machine = "EPISODE"
return None
# 2. Episode
if self.get_state_machine() == "EPISODE":
self.episodes += 1
# Add some boolean condition checks
solved = self.running_reward >= 40
capped = self.episodes > self.max_episodes
# Stop the episode
if (solved or capped):
self.episodes -= 1
print(self._get_training_status())
result = "Capped" if capped else "Solved"
print(f" -> Training loop result in a '{result}' state")
# Save our model
print(" -> Saving Model.")
self.save({
"model": self.model.to_json(),
"target": self.target.to_json(),
"actions:": self.action_history
})
print(" -> Cleaning Episode State.")
self.state_machine = "COMPLETE"
self.episodes = 0
print(" -> Training Complete.")
self.state_machine = "ONLINE"
return None
# Start the Episode
self.frames = 0
episode_reward = 1
# Update running reward to check condition for solving
self.episode_reward_history.append(episode_reward)
self.episode_reward_history = self.episode_reward_history[-100:]
self.running_reward = np.mean(self.episode_reward_history)
self.state_machine = "PRE_FRAME"
return None
# 3. Pre-Frame
if self.get_state_machine() == "PRE_FRAME":
# Reset if frames are finished
if self.frames >= self.max_frames:
self.state_machine = "EPISODE"
return None
# Execute a pre_frame. eg get_action
self.frames += 1
# Select Action to return
action = self._select_action(self.action_space)
self.queued_action = action
# e.g. {"type": "Belt", "x": 2, "y": y, "rotation": 270}
# action -> {"x", "y", "type", "rotation"}
print(self._get_training_status(), end="\r")
self.state_machine = "POST_FRAME"
return self.get_queued_action()
# X. -> Client runs between these
# 4. Post-Frame
if self.get_state_machine() == "POST_FRAME":
# Do something with the result of the frame
# 4. Validate the action by assigning a score.
# reward = self.validate(state)
# self.running_reward += reward
# Move to next frame.
self.state_machine = "PRE_FRAME"
return None
return None
def _step(self, game):
""" Advances the model one step.
Rhys Notes:
- DO I NEED TO RESET ALL HYPERPARAMETERS TO ORIGINAL VALUES AT START
OF EACH EPISODE????
- i think we do need a state because thats how tensorflow works to use
prebuilt methods
"""
# 1. Get state and action space from current GameState
state = self.region
# 2. Choose available action for frame
action = self._select_action(self.action_space)
self.queued_action = action
# 3. Apply the action
# state_next = game.step(self.queued_action)
return
# state_next = np.array(state_next)
# state_next, reward, goal = (state, 0, 0)
# 4. Validate the action by assigning a score.
reward = self.validate(state)
self.running_reward += reward
# 5. Log the events
self.log(state, state_next, action, reward, goal)
# state = state_next
# Update every fourth frame and once batch size is over 32
if (self.frames % self.update_after_actions == 0 and
len(self.goal_history) > self.batch_size):
# Get indices of samples for replay buffers
indices = np.random.choice(
range(len(self.goal_history)), size=self.batch_size
)
# Using list comprehension to sample from replay buffer
state_sample = np.array([self.state_history[i] for i in indices])
state_next_sample = np.array(
[self.state_next_history[i] for i in indices]
)
rewards_sample = [self.rewards_history[i] for i in indices]
action_sample = [self.action_history[i] for i in indices]
goal_sample = keras.ops.convert_to_tensor(
[float(self.goal_history[i]) for i in indices]
)
# Build the updated Q-values for the sampled future states
# Use the target model for stability
future_rewards = self.target.predict(state_next_sample)
# Q value = reward + discount factor * expected future reward
updated_q_values = rewards_sample + self.gamma * keras.ops.amax(
future_rewards, axis=1
)
# If final frame set the last value to -1
updated_q_values = updated_q_values*(1-goal_sample) - goal_sample
# Create a mask so we only calculate loss on the updated Q-values
masks = keras.ops.one_hot(action_sample, num_actions)
with tf.GradientTape() as tape:
# Train the model on the states and updated Q-values
q_values = self.model(state_sample)
# Apply the masks to the Q-values to get the Q-value for
# action taken
q_action = keras.ops.sum(
keras.ops.multiply(q_values, masks), axis=1
)
# Calculate loss between new Q-value and old Q-value
loss = self.loss_function(updated_q_values, q_action)
# Backpropagation
grads = tape.gradient(loss, self.model.trainable_variables)
self.optimiser.apply_gradients(
zip(grads, self.model.trainable_variables)
)
# -> Printing
if self.frames % self.update_target_network == 0:
# update the the target network with new weights
self.target.set_weights(self.model.get_weights())
# Log details
template = "running reward: {:.2f} at episode {}, frame count {}"
print(template.format(
self.running_reward,
self.episodes,
self.frames)
)
return
def _select_action(self, action_space):
""" Select an action using an epsilon-greedy strategy. """
action = None
# TODO Naive Random Implementation, fix for Epsilon Greedy
eps = np.random.random()
position = random.choice(list(action_space.keys()))
action = random.choice(action_space[position])
direction = random.choice([0, 90, 180, 270])
# Get direction from action
return {position: action}
# Take Random or attempt prediction.
if (self.frames < self.epsilon_random_frames or
self.epsilon > np.random.rand(1)[0]):
# Take random action
action = random.choice(action_space)
else:
# Predict action Q-Value from Environment
state_tensor = keras.ops.convert_to_tensor(state)
state_tensor = keras.ops.expand_dims(state_tensor, 0)
action_probs = self.model(state_tensor, training=False)
# TODO Adjust actions by weights
# Take best action
action = keras.ops.argmax(action_probs[0]).numpy()
# Decay probability of taking random action
self.epsilon -= self.epsilon_interval / self.epsilon_greedy_frames
self.epsilon = max(self.epsilon, self.epsilon_min)
return action
def get_queued_action(self):
""" Returns the queued action from the model and removes it. """
action = self.queued_action
self.queued_action = None
return action
# reward function -- very important for performance
# things to check for: (using random numbers)
# -- im scared to make rewards for non immediate goals so model does not find some hack
# - produce goal shape (+1)
# - produce future goal shape (+0.00001)
# - belts connecting (+0.0001)
# - belts connecting to hub (+0.0001)
# - plus more... idk, could add heps here dpends how complex we want this method to be
"""
gonna rewrite using ECS thing if possible???
should be easier then state i think
"""
def validate(self, level, entities, resources, products):
# level - current level of the game (used to get goal products)
# entities - list of buildings and respective locations
# resources - list of resources and respective locations
# products - list of products produced since last check
reward = 0
current_goal = GOALS[level] # need to make goals
# check if produced items are in current goal
for g in products:
if g in current_goal[0]: # only works for levels w 1 goal product
reward += 0.1
# check for miners on resources -- subgoal
# also finite amount, i suspect this will make model put miner on every available resource
for e in entities.keys():
if e in resources.keys(): # assume only miners can be placed on resources
reward += 0.01
# check for connected belts
# -- problem with making broader rewards is that model could find loophole and make a super long belt chain for example
# problem with not making these rewards, is that model will not know what is "good" until goal product is made, very slow
# could be some way to have these "sub-goals" only give reward up to some number of frames
# if frame < sub_goal_frame:
# allow subgoals
# alternatively, could use level or something
# if level > 2 ---> dont allow subgoal rewards
for pos in entities.keys():
if entities[pos] == "belt":
pass
# check belt connected to HUB
return reward
def log(self, state, state_next, action, reward, goal):
""" Updates the logged event history. """
# Update the logs
self.action_history.append(action)
self.state_history.append(state)
self.state_next_history.append(state_next)
self.goal_history.append(goal)
self.rewards_history.append(reward)
# Limit the state and reward history
if len(self.rewards_history) > self.max_memory_length:
del self.state_history[:1]
del self.state_next_history[:1]
del self.rewards_history[:1]
del self.goal_history[:1]
if __name__ == "__main__":
print("Please call this module as a dependency or import.")