def __init__(self, game):
super().__init__(name="SUB WATTEN BAGGING AGENT")
self.game = game
x, y = game.get_observation_size()
nnet1 = SubWattenNNet(x, y, 1, game.get_action_size())
nnet2 = AsymmetricSubWattenNNet(x, y, 1, game.get_action_size())
# Setting up the symmetric and asymmetric nnet
self.agent_symmetric = AgentNNet(nnet1)
self.agent_asymmetric = AgentNNet(nnet2)
try:
self.agent_symmetric.load("games/sub_watten/training/best.h5")
self.agent_asymmetric.load(
"games/asymmetric_sub_watten/training/v2/best.h5")
except OSError:
print("File not found with games/sub_watten/training/best.h5")
print(
"Maybe you are creating an agent for test purposes. I'll try to load the model from a different path"
)
self.agent_symmetric.load("../../sub_watten/training/best.h5")
self.agent_asymmetric.load(
"../../asymmetric_sub_watten/training/v2/best.h5")
def build_hand_watten_evaluate_agent(self,
agent_profile,
native_multi_gpu_enabled=False):
game = self.game_mapping[agent_profile.game]
agent_nnet = None
if agent_profile == EnvironmentSelector.HAND_WATTEN_EVALUATE:
x, y = game.get_observation_size()
nnet = HandWattenNNet(x, y, 1, game.get_action_size())
agent_nnet = AgentNNet(nnet)
elif agent_profile == EnvironmentSelector.HAND_WATTEN_EVALUATE_M_M:
x, y = game.get_observation_size()
nnet = MediumMediumNNet(x, y, 1, game.get_action_size())
agent_nnet = AgentNNet(nnet)
elif agent_profile == EnvironmentSelector.HAND_WATTEN_EVALUATE_S_S:
x, y = game.get_observation_size()
nnet = EasyEasyNNet(x, y, 1, game.get_action_size())
agent_nnet = AgentNNet(nnet)
elif agent_profile == EnvironmentSelector.HAND_WATTEN_EVALUATE_CNN:
x, y, z = game.get_observation_size()
nnet = CNNWatten(x, y, z, game.get_action_size())
agent_nnet = AgentNNet(nnet)
return agent_nnet
def build_hand_watten_train_agent(self,
agent_profile,
native_multi_gpu_enabled=False):
game = self.game_mapping[agent_profile.game]
if agent_profile == EnvironmentSelector.HAND_WATTEN_TRAIN:
x, y = game.get_observation_size()
nnet = HandWattenNNet(x, y, 1, game.get_action_size())
agent_nnet = AgentNNet(nnet)
print("Configuring build_hand_watten_train_agent...")
return AgentMCTS(agent_nnet,
exp_rate=AgentMCTS.EXPLORATION_RATE_MEDIUM,
numMCTSSims=30,
max_predict_time=10,
num_threads=1)
elif agent_profile == EnvironmentSelector.HAND_WATTEN_TRAIN_S_S:
x, y = game.get_observation_size()
nnet = EasyEasyNNet(x, y, 1, game.get_action_size())
agent_nnet = AgentNNet(nnet)
print("Configuring build_hand_watten_s_s...")
return AgentMCTS(agent_nnet,
exp_rate=AgentMCTS.EXPLORATION_RATE_MEDIUM,
numMCTSSims=30,
max_predict_time=10,
num_threads=1)
elif agent_profile == EnvironmentSelector.HAND_WATTEN_TRAIN_M_M:
x, y = game.get_observation_size()
nnet = MediumMediumNNet(x, y, 1, game.get_action_size())
agent_nnet = AgentNNet(nnet)
print("Configuring build_hand_watten_m_m...")
return AgentMCTS(agent_nnet,
exp_rate=AgentMCTS.EXPLORATION_RATE_MEDIUM,
numMCTSSims=30,
max_predict_time=10,
num_threads=1)
elif agent_profile == EnvironmentSelector.HAND_WATTEN_TRAIN_CNN:
x, y, z = game.get_observation_size()
nnet = CNNWatten(x, y, z, game.get_action_size())
agent_nnet = AgentNNet(nnet)
print("Configuring build hand watten CNN....")
return AgentMCTS(agent_nnet,
exp_rate=AgentMCTS.EXPLORATION_RATE_MEDIUM,
numMCTSSims=30,
max_predict_time=10,
num_threads=1)
return None
def build_watten_train_4_512_agent(self,
agent_profile,
native_multi_gpu_enabled=False):
game = self.game_mapping[agent_profile.game]
x, y = game.get_observation_size()
nnet = WattenNNet4x512(x, y, 1, game.get_action_size())
agent_nnet = AgentNNet(nnet)
if agent_profile == EnvironmentSelector.WATTEN_AGENT_4_512_TRAIN:
print("Configuring build_watten_train_4_512_agent...")
return AgentMCTS(agent_nnet,
exp_rate=AgentMCTS.EXPLORATION_RATE_MEDIUM,
numMCTSSims=100,
max_predict_time=10,
num_threads=16)
elif agent_profile == EnvironmentSelector.WATTEN_AGENT_4_512_EVALUATE:
print("Configuring build_watten_evaluate_4_512_agent...")
return AgentMCTS(agent_nnet,
exp_rate=AgentMCTS.EXPLORATION_RATE_MEDIUM,
numMCTSSims=30,
max_predict_time=10,
num_threads=16,
name="build_watten_evaluate_4_512_agent")
elif agent_profile == EnvironmentSelector.WATTEN_AGENT_HUMAN:
return WattenHumanAgent(game)
elif agent_profile == EnvironmentSelector.WATTEN_AGENT_NNET:
agent_nnet.load(
os.path.abspath(
"../games/watten/training/best-4-512-new-4.h5"))
return agent_nnet
return None
def build_watten_train_big_agent(self,
agent_profile,
native_multi_gpu_enabled=False):
game = self.game_mapping[agent_profile.game]
x, y = game.get_observation_size()
nnet = WattenNNetFirstLayerBig(x, y, 1, game.get_action_size())
agent_nnet = AgentNNet(nnet)
if agent_profile == EnvironmentSelector.WATTEN_AGENT_BIG_TRAIN:
print("Configuring build_watten_train_big_agent...")
return AgentMCTS(agent_nnet,
exp_rate=AgentMCTS.EXPLORATION_RATE_MEDIUM,
numMCTSSims=100,
max_predict_time=10,
num_threads=16)
elif agent_profile == EnvironmentSelector.WATTEN_AGENT_BIG_EVALUATE:
print("Configuring build_watten_evaluate_big_agent...")
return AgentMCTS(agent_nnet,
exp_rate=AgentMCTS.EXPLORATION_RATE_MEDIUM,
numMCTSSims=2,
max_predict_time=10,
num_threads=16,
name="build_watten_evaluate_big_agent")
elif agent_profile == EnvironmentSelector.WATTEN_AGENT_HUMAN:
return WattenHumanAgent(game)
return None
def build_sub_watten_train_agent(self,
agent_profile,
native_multi_gpu_enabled=False):
game = self.game_mapping[agent_profile.game]
x, y = game.get_observation_size()
if agent_profile == EnvironmentSelector.SUB_WATTEN_AGENT_TRAIN:
nnet = SubWattenNNet(x, y, 1, game.get_action_size())
elif agent_profile == EnvironmentSelector.SUB_WATTEN_AGENT_TRAIN_SIMPLE:
nnet = SubWattenSimplerNNet(x, y, 1, game.get_action_size())
agent_nnet = AgentNNet(nnet)
if agent_profile == EnvironmentSelector.SUB_WATTEN_AGENT_TRAIN or \
agent_profile == EnvironmentSelector.SUB_WATTEN_AGENT_TRAIN_SIMPLE:
print("Configuring build_sub_watten_train_agent...")
return AgentMCTS(agent_nnet,
exp_rate=AgentMCTS.EXPLORATION_RATE_MEDIUM,
numMCTSSims=100,
max_predict_time=10,
num_threads=1)
return None
def build_horovod_checkers_agent(self,
agent_profile,
native_multi_gpu_enabled=False):
from games.checkers.nnet.CheckersResNNetDistributed import CheckersResNNetDistributed
assert not native_multi_gpu_enabled, "ERROR: Horovod NNet does not support native multi-gpu mode!"
game = self.game_mapping[agent_profile.game]
nnet = CheckersResNNetDistributed(game.get_observation_size()[0],
game.get_observation_size()[1],
game.get_observation_size()[2],
game.get_action_size(),
horovod_distributed=True)
agent_nnet = AgentNNet(nnet)
if agent_profile == EnvironmentSelector.CHECKERS_AGENT_TRAIN_RCNN_DISTRIBUTED:
return AgentMCTS(agent_nnet,
exp_rate=AgentMCTS.EXPLORATION_RATE_INIT,
numMCTSSims=1500,
max_predict_time=5)
elif agent_profile == EnvironmentSelector.CHECKERS_AGENT_TEST_AGENT_RCNN_DISTRIBUTED:
return AgentMCTS(agent_nnet,
exp_rate=AgentMCTS.NO_EXPLORATION,
numMCTSSims=1500,
max_predict_time=10)
else:
return None
def build_tpu_checkers_agent(self,
agent_profile,
native_multi_gpu_enabled=False):
from games.checkers.nnet.CheckersResNNetTPU import CheckersResNNetTPU
assert not native_multi_gpu_enabled, "ERROR: TPU NNet does not support native multi-gpu mode!"
game = self.game_mapping[agent_profile.game]
nnet = CheckersResNNetTPU(game.get_observation_size()[0],
game.get_observation_size()[1],
game.get_observation_size()[2],
game.get_action_size())
agent_nnet = AgentNNet(nnet)
if agent_profile == EnvironmentSelector.CHECKERS_AGENT_TRAIN_RCNN_TPU:
return AgentMCTS(agent_nnet,
exp_rate=AgentMCTS.EXPLORATION_RATE_INIT,
numMCTSSims=200,
max_predict_time=None,
verbose=False,
num_threads=1)
else:
return None
def build_native_checkers_rcnn_agent(self,
agent_profile,
native_multi_gpu_enabled=False):
game = self.game_mapping[agent_profile.game]
if not native_multi_gpu_enabled:
nnet = CheckersResNNet(game.get_observation_size()[0],
game.get_observation_size()[1],
game.get_observation_size()[2],
game.get_action_size())
else:
nnet = CheckersResNNet(game.get_observation_size()[0],
game.get_observation_size()[1],
game.get_observation_size()[2],
game.get_action_size(),
multi_gpu=True,
multi_gpu_n=len(GPUtil.getGPUs()))
agent_nnet = AgentNNet(nnet)
if agent_profile == EnvironmentSelector.CHECKERS_AGENT_TRAIN_RCNN_DEFAULT:
return AgentMCTS(agent_nnet,
exp_rate=AgentMCTS.EXPLORATION_RATE_INIT,
numMCTSSims=1500,
max_predict_time=3,
num_threads=1)
elif agent_profile == EnvironmentSelector.CHECKERS_AGENT_TEST_AGENT_RCNN_DEFAULT:
return AgentMCTS(agent_nnet,
exp_rate=AgentMCTS.NO_EXPLORATION,
numMCTSSims=1500,
max_predict_time=10,
num_threads=2,
verbose=True)
else:
return None
def build_asymmetric_sub_watten_evaluate_agent(
self, agent_profile, native_multi_gpu_enabled=False):
game = self.game_mapping[agent_profile.game]
x, y = game.get_observation_size()
nnet = AsymmetricSubWattenNNet(x, y, 1, game.get_action_size())
agent_nnet = AgentNNet(nnet)
return agent_nnet
def build_sub_watten_evaluate_agent(self,
agent_profile,
native_multi_gpu_enabled=False):
game = self.game_mapping[agent_profile.game]
x, y = game.get_observation_size()
if agent_profile == EnvironmentSelector.SUB_WATTEN_AGENT_EVALUATE:
nnet = SubWattenNNet(x, y, 1, game.get_action_size())
elif agent_profile == EnvironmentSelector.SUB_WATTEN_AGENT_EVALUATE_SIMPLE:
nnet = SubWattenSimplerNNet(x, y, 1, game.get_action_size())
agent_nnet = AgentNNet(nnet)
return agent_nnet
def build_durak_train_agent(self,
agent_profile,
native_multi_gpu_enabled=False):
game = self.game_mapping[agent_profile.game]
x, y = game.get_observation_size()
nnet = DurakNNet(x, y, 1, game.get_action_size())
agent_nnet = AgentNNet(nnet)
if agent_profile == EnvironmentSelector.DURAK_AGENT_TRAIN:
return AgentMCTS(agent_nnet,
exp_rate=AgentMCTS.EXPLORATION_RATE_MEDIUM,
numMCTSSims=100,
max_predict_time=10)
return None
def build_tictactoe_train_agent(self,
agent_profile,
native_multi_gpu_enabled=False):
game = self.game_mapping[agent_profile.game]
nnet = TicTacToeNNet(game.get_observation_size()[0],
game.get_observation_size()[1],
game.get_observation_size()[2],
game.get_action_size())
agent_nnet = AgentNNet(nnet)
if agent_profile == EnvironmentSelector.TICTACTOE_AGENT_TRAIN:
return AgentMCTS(agent_nnet,
exp_rate=AgentMCTS.EXPLORATION_RATE_MEDIUM,
numMCTSSims=100,
max_predict_time=10)
return None
def test_nn_agent_prediction(self):
sub_watten_game = WattenSubGame()
clone_sub_watten_game = sub_watten_game.clone()
x, y = sub_watten_game.get_observation_size()
nnet = SubWattenNNet(x, y, 1, sub_watten_game.get_action_size())
agent_nnet = AgentNNet(nnet)
agent_nnet.load("../../sub_watten/training/best.h5")
pi_values, v = agent_nnet.predict(sub_watten_game,
sub_watten_game.get_cur_player())
clone_pi_values, clone_v = agent_nnet.predict(
clone_sub_watten_game, clone_sub_watten_game.get_cur_player())
self.assertEqual(pi_values.all(), clone_pi_values.all())
self.assertEqual(v, clone_v)
def sub_watten_non_human_agent_for_total_watten(self):
game = WattenSubGame()
x, y = game.get_observation_size()
nnet = SubWattenNNet(x, y, 1, game.get_action_size())
agent_nnet = AgentNNet(nnet)
print('Building sub_watten non human agent for total_watten')
# load here best sub_watten model
try:
agent_nnet.load("games/sub_watten/training/default_nn/best.h5")
except OSError:
print("File not found with games/sub_watten/training/best.h5")
print(
"Maybe you are creating an agent for test purposes. I'll try to load the model from a different path"
)
agent_nnet.load("../../sub_watten/training/default_nn/best.h5")
return agent_nnet