-
Notifications
You must be signed in to change notification settings - Fork 16
/
Copy pathwheatley.py
214 lines (165 loc) · 9.43 KB
/
wheatley.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
#
# Wheatley
# Copyright (c) 2023 Jolibrain
# Authors:
# Guillaume Infantes <[email protected]>
# Antoine Jacquet <[email protected]>
# Michel Thomazo <[email protected]>
# Emmanuel Benazera <[email protected]>
#
#
# This file is part of Wheatley.
#
# Wheatley is free software: you can redistribute it and/or modify
# it under the terms of the GNU General Public License as published by
# the Free Software Foundation, either version 3 of the License, or
# (at your option) any later version.
#
# Wheatley is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
# GNU General Public License for more details.
#
# You should have received a copy of the GNU General Public License
# along with Wheatley. If not, see <https://www.gnu.org/licenses/>.
#
import datetime
import pathlib
import gym
import numpy
import torch
from muzero_general.games.abstract_game import AbstractGame
from env.env import Env
class MuZeroConfig:
def __init__(self):
self.problem_description = None
self.env_specification = None
self.ortools_ratio = 0.5
self.test_policy = True
# fmt: off
# More information is available here: https://github.com/werner-duvaud/muzero-general/wiki/Hyperparameter-Optimization
self.seed = 0 # Seed for numpy, torch and the game
self.max_num_gpus = None # Fix the maximum number of GPUs to use. It's usually faster to use a single GPU (set it to 1) if it has enough memory. None will use every GPUs available
### Game
#self.observation_shape = (1, 1, 4) # Dimensions of the game observation, must be 3D (channel, height, width). For a 1D array, please reshape it to (1, 1, length of array)
#self.action_space = list(range(2)) # Fixed list of all possible actions. You should only edit the length
self.players = list(range(1)) # List of players. You should only edit the length
self.stacked_observations = 0 # Number of previous observations and previous actions to add to the current observation
# Evaluate
self.muzero_player = 0 # Turn Muzero begins to play (0: MuZero plays first, 1: MuZero plays second)
self.opponent = None # Hard coded agent that MuZero faces to assess his progress in multiplayer games. It doesn't influence training. None, "random" or "expert" if implemented in the Game class
### Self-Play
self.num_workers = 1 # Number of simultaneous threads/workers self-playing to feed the replay buffer
self.selfplay_on_gpu = False
self.max_moves = None # Maximum number of moves if game is not finished before
self.num_simulations = 16 # Number of future moves self-simulated
self.discount = 1 # Chronological discount of the reward
self.temperature_threshold = None # Number of moves before dropping the temperature given by visit_softmax_temperature_fn to 0 (ie selecting the best action). If None, visit_softmax_temperature_fn is used every time
# Root prior exploration noise
self.root_dirichlet_alpha = 0.25
self.root_exploration_fraction = 0.25
# UCB formula
self.pb_c_base = 19652
self.pb_c_init = 1.25
### Network
self.network = "fullyconnected" # "resnet" / "fullyconnected"
self.support_size = 10 # Value and reward are scaled (with almost sqrt) and encoded on a vector with a range of -support_size to support_size. Choose it so that support_size <= sqrt(max(abs(discounted reward)))
# Residual Network
self.downsample = False # Downsample observations before representation network, False / "CNN" (lighter) / "resnet" (See paper appendix Network Architecture)
self.blocks = 1 # Number of blocks in the ResNet
self.channels = 2 # Number of channels in the ResNet
self.reduced_channels_reward = 2 # Number of channels in reward head
self.reduced_channels_value = 2 # Number of channels in value head
self.reduced_channels_policy = 2 # Number of channels in policy head
self.resnet_fc_reward_layers = [] # Define the hidden layers in the reward head of the dynamic network
self.resnet_fc_value_layers = [] # Define the hidden layers in the value head of the prediction network
self.resnet_fc_policy_layers = [] # Define the hidden layers in the policy head of the prediction network
# Fully Connected Network
self.encoding_size = 128
self.fc_representation_layers = [128] # Define the hidden layers in the representation network
self.fc_dynamics_layers = [128] # Define the hidden layers in the dynamics network
self.fc_reward_layers = [128] # Define the hidden layers in the reward network
self.fc_value_layers = [128] # Define the hidden layers in the value network
self.fc_policy_layers = [128] # Define the hidden layers in the policy network
### Training
self.results_path = pathlib.Path(__file__).resolve().parents[1] / "results" / pathlib.Path(__file__).stem / datetime.datetime.now().strftime("%Y-%m-%d--%H-%M-%S") # Path to store the model weights and TensorBoard logs
self.save_model = True # Save the checkpoint in results_path as model.checkpoint
self.training_steps = 1000000 # Total number of training steps (ie weights update according to a batch)
self.batch_size = 32 # Number of parts of games to train on at each training step
self.checkpoint_interval = 10 # Number of training steps before using the model for self-playing
self.value_loss_weight = 1 # Scale the value loss to avoid overfitting of the value function, paper recommends 0.25 (See paper appendix Reanalyze)
self.train_on_gpu = torch.cuda.is_available() # Train on GPU if available
self.optimizer = "Adam" # "Adam" or "SGD". Paper uses SGD
self.weight_decay = 1e-4 # L2 weights regularization
self.momentum = 0.9 # Used only if optimizer is SGD
# Exponential learning rate schedule
self.lr_init = 2e-5 # Initial learning rate
self.lr_decay_rate = 1 # Set it to 1 to use a constant learning rate
self.lr_decay_steps = 10000
### Replay Buffer
self.replay_buffer_size = 100 # Number of self-play games to keep in the replay buffer
self.num_unroll_steps = 5 # Number of game moves to keep for every batch element
self.td_steps = None # Number of steps in the future to take into account for calculating the target value
self.PER = False # Prioritized Replay (See paper appendix Training), select in priority the elements in the replay buffer which are unexpected for the network
self.PER_alpha = 0.5 # How much prioritization is used, 0 corresponding to the uniform case, paper suggests 1
# Reanalyze (See paper appendix Reanalyse)
self.use_last_model_value = False # Use the last model to provide a fresher, stable n-step value (See paper appendix Reanalyze)
self.reanalyse_on_gpu = False
### Adjust the self play / training ratio to avoid over/underfitting
self.self_play_delay = 0 # Number of seconds to wait after each played game
self.training_delay = 0 # Number of seconds to wait after each training step
self.ratio = None # Desired training steps per self played step ratio. Equivalent to a synchronous version, training can take much longer. Set it to None to disable it
# fmt: on
def visit_softmax_temperature_fn(self, trained_steps):
"""
Parameter to alter the visit count distribution to ensure that the action selection becomes greedier as training progresses.
The smaller it is, the more likely the best action (ie with the highest visit count) is chosen.
Returns:
Positive float.
"""
return 1
class Game(AbstractGame):
"""
Game wrapper.
"""
def __init__(self, seed, config):
self.env = Env(config.problem_description, config.env_specification)
def step(self, action):
"""
Apply action to the game.
Args:
action : action of the action_space to take.
Returns:
The new observation, the reward and a boolean if the game has ended.
"""
observation, reward, done, _ = self.env.step(action)
return observation, reward, done
def legal_actions(self):
"""
Should return the legal actions at each turn, if it is not available, it can return
the whole action space. At each turn, the game have to be able to handle one of returned actions.
For complex game where calculating legal moves is too long, the idea is to define the legal actions
equal to the action space but to return a negative reward if the action is illegal.
Returns:
An array of integers, subset of the action space.
"""
mask = self.env.action_masks()
legals = list(numpy.nonzero(mask)[0])
return legals
def reset(self, soft=False):
"""
Reset the game for a new game.
Returns:
Initial observation of the game.
"""
return self.env.reset(soft=soft)
def close(self):
"""
Properly close the game.
"""
pass
def render(self):
"""
Display the game observation.
"""
pass