-
Notifications
You must be signed in to change notification settings - Fork 0
/
dqn.py
232 lines (174 loc) · 14.6 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
import os
import sys
import time
import random
from itertools import combinations
import numpy as np
import tensorflow as tf
from tensorflow.keras.models import Sequential,clone_model,save_model,load_model
from tensorflow.keras.layers import Dense
from tensorflow.keras.optimizers import Adam
sys.path.insert(0,'games') # to ensure game modules import correctly
########################################################################################################
########################################## LEARNING ALGORITHM ##########################################
########################################################################################################
class RandomNetwork: # "network" that plays only random moves
def __init__(self,game): self.game = game
def __call__(self,state):
action = self.game.sample_legal_move()
return np.array( [ [ 1 if action == i else 0 for i in range(game.get_action_space_size()) ] ] )
def make_q_network(game,config): # make a neural network according to config parameters
network_layers = [ Dense(config['network']['num_hidden_units'],activation=config['network']['hidden_activation'],
kernel_initializer='glorot_normal',input_shape=(game.get_observation_shape(),)) ]
for _ in range(config['network']['num_hidden_layers']-1): network_layers.append( Dense(config['network']['num_hidden_units'],activation=config['network']['hidden_activation']) )
network_layers.append( Dense(game.get_action_space_size(),activation=config['network']['output_activation'],kernel_initializer='glorot_normal') )
q_network = Sequential(network_layers)
q_network.compile(loss=config['network']['loss'],optimizer=Adam(learning_rate=config['network']['learning_rate']))
return q_network
def q_learning(game,config):
# q_network_history stores previous network iterations, to be used when testing current network parameters against past parameters
q_network_history = { 0:RandomNetwork(game) } # initialize with a random network at episode 0 to serve as a baseline when comparing future network performances
q_network = make_q_network(game,config)
replay_memory = [] # list[ tuple(state_t,action_t,reward_(t+1)) ]
start_time = time.time()
for episode_num in range(1,int(config['self_play']['num_episodes'])+1):
game.reset()
state_histories = [ [] for _ in range(game.get_number_of_players()) ] # list[ state_history_player0_list[state] , state_history_player1_list[state] , ... ]
action_histories = [ [] for _ in range(game.get_number_of_players()) ] # list[ action_history_player0_list[action] , action_history_player0_list[action] , ... ]
state = game.get_features( game.get_player_turn() )
while True:
player_turn = game.get_player_turn()
if np.random.sample() < config['self_play']['epsilon']: action = game.sample_legal_move()
else: action = np.argmax( q_network(state)[0,:] )
legal = game.act(action)
new_state = game.get_features( game.get_player_turn() ) # in the perspective of opponent
done,results_dict = game.check_game_over() # results_dict is a dict{ player_index:'win'/'loss'/'tie' , ... }
state_histories[player_turn].append(state)
action_histories[player_turn].append(action)
if done or (not legal): break
state = new_state
# only append the most recent state and action to the replay buffer if the episode terminated due to an illegal action
if not legal: replay_memory.append( ( state_histories[player_turn][-1] , action_histories[player_turn][-1] , config['reward']['illegal_move'] ) )
elif done: # otherwise, if the episode terminated naturally with a game over, calculate discounted rewards from the terminal reward and append all states and actions encountered to the replay buffer
for player_index in range(game.get_number_of_players()):
reward = config['reward'][results_dict[player_index]]
expected_return = reward
for i in range(len(state_histories[player_index])-1,-1,-1):
replay_memory.append( ( state_histories[player_index][i] , action_histories[player_index][i] , expected_return ) )
expected_return *= config['reward']['discount_factor']
for _ in range( max( 0 , len(replay_memory) - config['training']['replay_memory_capacity'] ) ): replay_memory.pop(0) # remove the oldest replay samples if replay buffer is over capacity
if episode_num % config['training']['train_interval'] == 0 and len(replay_memory) > config['training']['batch_size']: # train the network every set interval denoted by config
training_batch = np.vstack( random.sample(replay_memory,config['training']['batch_size']) )
state_batch , action_batch , reward_batch = np.vstack(training_batch[:,0]) , training_batch[:,1] , training_batch[:,2]
# state_batch is a (batch_size x game_observation_shape) matrix , action_batch and reward_batch are (batch_size,) vectors
reward_matrix = q_network(state_batch).numpy() # reward_matrix is a (batch_size x game_action_space_size) matrix, where each entry is the corresponding action value
reward_matrix[ range(reward_matrix.shape[0]) , action_batch.astype(int) ] = reward_batch # assign the corresponding actions taken in the sample trajectories with the associated discounted reward
q_network.fit(state_batch, reward_matrix, epochs=1, verbose=0)
if episode_num % config['testing']['test_interval'] == 0:
print(f'\nIteration {episode_num} complete - {time.time()-start_time} seconds')
ordered_model_keys = list(q_network_history.keys()) # order the q_network episode_num keys in ascending order
ordered_model_keys.sort()
# test against all combinations of previous models against the current network's parameters
for model_name_list in combinations( ordered_model_keys , game.get_number_of_players()-1 ):
model_list = [q_network_history[model_name] for model_name in model_name_list] + [q_network] # for every possible combination of opponents, add our current q_network to the model_list
test_ai( game , config['testing']['num_test_episodes'] , list(model_name_list) + [episode_num] , model_list ) # append current episode_num key to model_name_list and test games with the models in model_list
if not os.path.exists(f'models/{game.get_name()}'): os.mkdir(f'models/{game.get_name()}') # create a subfolder for this game under the 'models' directory
save_model(q_network,f'models/{game.get_name()}/{episode_num}.h5') # save the current model
saved_network = clone_model(q_network) # add the current model to q_network_history
saved_network.set_weights(q_network.get_weights())
q_network_history[episode_num] = saved_network
start_time = time.time()
#######################################################################################################
########################################## UTILITY FUNCTIONS ##########################################
#######################################################################################################
def play(game,model_name_list,model_list=[],print_game=True): # play the game against human(s) and/or q_network(s)
# the number of human players is equal to ( game.get_number_of_players() - len(model_name_list) ), so you could play anything from an all human game, all the way to an all AI game
# model_name_list: list[ model_name_strings ]
# model_list: list[ keras_models_corresponding_to_model_name_strings_in_model_name_list ], if this is empty, then manually load models in using model filenames from model_name_list
q_networks = {} # dict{ player_index:model }
q_network_names = {} # dict{ player_index:model_name_string }
player_indices = [i for i in range(game.get_number_of_players())] # initialize with all possible player indices, and then randomly take some away and assign them to AI models
for i,model_name in enumerate(model_name_list): # randomly assign player indices for AI players
player_index = np.random.choice(player_indices,1)[0] #i
if len(model_list) > 0: q_networks[player_index] = model_list[i] # if model_list isn't empty, that means the models have already been instantiated
else: q_networks[player_index] = load_model(f"models/{game.get_name()}/{str(model_name)+'.h5'}") # otherwise instantiate the models
q_network_names[player_index] = model_name
player_indices.remove(player_index)
if print_game: print(f'Player indices:{player_indices}')
if print_game: print(f'AI model player indices:{list(q_networks.keys())}')
# begin self play
game.reset()
state = game.get_features( game.get_player_turn() )
while True:
if print_game: print(game)
player_index = game.get_player_turn()
if player_index in player_indices: action = int( input('Move: ') ) # human move
else:
action = np.argmax( q_networks[player_index](state)[0,:] ) # AI move
if print_game: print(f'==== AI model move: {action} ====')
legal = game.act(action)
if not legal: return {player_index:'illegal'},q_network_names
done,results_dict = game.check_game_over()
if done: break
state = game.get_features( game.get_player_turn() )
if print_game: print(game)
return results_dict,q_network_names
def test_ai(game,n,model_name_list=[],model_list=[]): # test AI models against each other
# model_name_list: list[ model_name_strings ]
# model_list: list[ keras_models_corresponding_to_model_name_strings_in_model_name_list ]
# for every model in model_name_list/model_list, have them play n games against each other and compile winrate statistics
# every combination of models from model_name_list/model_list are tested once per iteration of n
# if model_name_list/model_list are empty, then load all models from the corresponding model directory for this game, and have them play against each other n times and compile winrate statistics
if model_name_list == []: models_dict = { model_filename.split('.h5')[0] : load_model(f'models/{game.get_name()}/{model_filename}') for model_filename in os.listdir(f'models/{game.get_name()}') } # dict{ model_name:model }
else: models_dict = { model_name_list[i] : model_list[i] for i in range(len(model_name_list)) }
stats = {} # dict { model_name: {'win':int,'tie':int,'loss':int} }
# get a list that contains all possible matchup combinations between all model AI's for this game
model_combinations = [ model_name_list_subset for model_name_list_subset in combinations( models_dict.keys() , game.get_number_of_players() ) ]
start_time = time.time()
for _ in range(n):
for model_name_list_subset in model_combinations: # model_name_list_subset is a subset of models to be tested in a game
model_list_subset = [models_dict[model_name] for model_name in model_name_list_subset] # list of actual keras models, corresponding to model_name_list_subset
results_dict,q_network_names = play(game,model_name_list_subset,model_list_subset,print_game=False) # play the game
for player_index in q_network_names: # update the result of the game for every model that participated
model_name = q_network_names[player_index] # player_indices are randomly assigned in game, so we need q_network_names to find the actual model_name
if model_name not in stats: stats[model_name] = {'win':0,'tie':0,'loss':0,'illegal':0}
# not all player indices will always be in results_dict if one player made an illegal move (in which case results_dict would only contain the index of the player that made that illegal move)
if player_index in results_dict: stats[model_name][results_dict[player_index]] += 1
print(f'Testing {n} games for models {list(models_dict.keys())} - {time.time()-start_time} seconds')
ordered_model_name_list = [] # keep track of model names
win_rate_list = [] # indices correspond to model indices in ordered_model_name_list
for model_name in models_dict.keys(): # populate the two lists
ordered_model_name_list.append(model_name)
win_rate_list.append( stats[model_name]['win'] / max( 1 , (stats[model_name]['win']+stats[model_name]['tie']+stats[model_name]['loss']+stats[model_name]['illegal']) ) ) # max to prevent 0 division error
for index in np.argsort(win_rate_list)[::-1]: # print results for each model in descending order according to winrate
model_name = ordered_model_name_list[index]
win,tie,loss,illegal = stats[model_name]['win'],stats[model_name]['tie'],stats[model_name]['loss'],stats[model_name]['illegal']
total = win+tie+loss+illegal
print(f'{model_name}\tWin:{win}/{total} ({round(100*win/max(1,total),2)}%)\tTie:{tie}/{total} ({round(100*tie/max(1,total),2)}%)\tLoss:{loss}/{total} ({round(100*loss/max(1,total),2)}%)\tIllegal:{illegal}/{total} ({round(100*illegal/max(1,total),2)}%)')
if __name__ == '__main__':
config = { 'self_play': {'num_episodes':1e10,
'epsilon':0.1},
'reward': {'win':1,
'tie':0,
'loss':-1,
'illegal_move':-1,
'discount_factor':1.},
'network': {'num_hidden_units':200,
'num_hidden_layers':3,
'hidden_activation':'relu',
'output_activation':'tanh',
'loss':'mse',
'learning_rate':1e-4},
'training': {'replay_memory_capacity':10000,
'batch_size':1000,
'train_interval':10},
'testing': {'test_interval':100000,
'num_test_episodes':10}
}
from GwentLite import GwentLite
game = GwentLite()
## play(game,[700000])
## with tf.device('/CPU:0'):
## q_learning(game,config)
## with tf.device('/CPU:0'):
## test_ai(game,11446,model_name_list=[],model_list=[])