def declare_trainer():
env = SC2Env(map_name="Simple64",
render=False,
step_mul=16,
player_race="terran",
enemy_race="terran",
difficulty="very_easy")
action_wrapper = SimpleTerranWrapper(atk_grid_x=4, atk_grid_y=4)
state_builder = MultipleUnitGridState(4)
helper = ModelBuilder()
helper.add_input_layer()
helper.add_fullyconn_layer(nodes=50)
helper.add_output_layer()
dq_network = DoubleDeepQLearning(action_wrapper=action_wrapper,
state_builder=state_builder,
build_model=helper.get_model_layout(),
use_memory=True,
gamma=0.99,
learning_rate=0.001,
memory_maxlen=100000,
min_memory_size=2000,
lib="keras",
epsilon_decay=0.9994,
epsilon_start=0.6,
epsilon_min=0.005,
epsilon_linear_decay=True,
per_episode_epsilon_decay=True)
agent = SC2Agent(dq_network, KilledUnitsReward())
# trainer = Trainer(env, agent, save_path='/home/lpdcalves/', file_name="tvt_veasy_newactwrapper_t1",
# save_every=200, enable_save=True, relative_path=False, reset_epsilon=False,
# max_training_episodes=3000, max_steps_training=1500,
# max_test_episodes=100, max_steps_testing=1500, rolling_avg_window_size=50)
trainer = Trainer(env,
agent,
save_path='urnai/models/saved',
file_name="terran_ddql_newgridstate7",
save_every=6,
enable_save=True,
relative_path=True,
reset_epsilon=False,
max_training_episodes=2,
max_steps_training=1000,
max_test_episodes=2,
max_steps_testing=100,
rolling_avg_window_size=5)
return trainer
def declare_trainer():
env = GymEnv(id="Breakout-ram-v0", render=True)
action_wrapper = env.get_action_wrapper()
#state_builder = PureState(env.env_instance.observation_space)
state_builder = GymState(env.env_instance.observation_space.shape[0])
helper = ModelBuilder()
# helper.add_input_layer(nodes=32)
# helper.add_fullyconn_layer(8)
helper.add_input_layer(nodes=66)
helper.add_output_layer()
# dq_network = DQNKeras(action_wrapper=action_wrapper, state_builder=state_builder,
# gamma=0.99, learning_rate=0.001, epsilon_decay=0.9995, epsilon_min=0.01,
# build_model=helper.get_model_layout(), memory_maxlen=5000)
# dq_network = DDQNKeras(action_wrapper=action_wrapper, state_builder=state_builder, build_model=helper.get_model_layout(), use_memory=False,
# gamma=0.99, learning_rate=0.001, epsilon_decay=0.999997, epsilon_min=0.01, memory_maxlen=100000, min_memory_size=2000)
dq_network = DeepQLearning(action_wrapper=action_wrapper,
state_builder=state_builder,
build_model=helper.get_model_layout(),
gamma=0.99,
learning_rate=0.01,
epsilon_decay=0.998,
epsilon_min=0.01,
memory_maxlen=100000,
min_memory_size=4000,
per_episode_epsilon_decay=True,
lib="pytorch")
#dq_network = PGKeras(action_wrapper, state_builder, learning_rate=0.001, gamma=0.99, build_model=helper.get_model_layout())
agent = GenericAgent(dq_network, PureReward())
trainer = Trainer(env,
agent,
save_path='urnai/models/saved',
file_name="breakout-ram-v0_dql_66_pytorch",
save_every=300,
enable_save=True,
relative_path=True,
max_training_episodes=1200,
max_steps_training=1800,
max_test_episodes=100,
max_steps_testing=1800)
return trainer
def declare_trainer():
env = GymEnv(id="CartPole-v1", render=False)
action_wrapper = env.get_action_wrapper()
state_builder = GymState(env.env_instance.observation_space.shape[0])
helper = ModelBuilder()
helper.add_input_layer(nodes=50)
#helper.add_fullyconn_layer(50)
helper.add_output_layer()
dq_network = DeepQLearning(action_wrapper=action_wrapper,
state_builder=state_builder,
build_model=helper.get_model_layout(),
gamma=0.99,
use_memory=True,
learning_rate=0.001,
epsilon_decay=0.9997,
epsilon_min=0.01,
memory_maxlen=50000,
min_memory_size=100,
batch_size=32,
lib="pytorch")
# dq_network = DQNPytorch(action_wrapper=action_wrapper, state_builder=state_builder, build_model=helper.get_model_layout(),
# gamma=0.99, learning_rate=0.001, epsilon_decay=0.9997, epsilon_min=0.01, memory_maxlen=50000, min_memory_size=100)
# dq_network = DDQNKeras(action_wrapper=action_wrapper, state_builder=state_builder, build_model=helper.get_model_layout(), use_memory=False,
# gamma=0.99, learning_rate=0.001, epsilon_decay=0.999997, epsilon_min=0.01, memory_maxlen=100000, min_memory_size=2000)
agent = GenericAgent(dq_network, PureReward())
# Cartpole-v1 is solved when avg. reward over 100 episodes is greater than or equal to 475
trainer = Trainer(env,
agent,
save_path='urnai/models/saved',
file_name="cartpole_v1_dql_pytorch1",
save_every=100,
enable_save=True,
relative_path=True,
max_training_episodes=1000,
max_steps_training=1000,
max_test_episodes=100,
max_steps_testing=1000)
return trainer
def declare_trainer():
env = GymEnv(id="Breakout-v0", render=True)
img = env.reset()
action_wrapper = env.get_action_wrapper()
state_builder = GymState(env.env_instance.observation_space.shape)
helper = ModelBuilder()
helper.add_convolutional_layer(
filters=8,
kernel_size=4,
input_shape=env.env_instance.observation_space.shape)
helper.add_maxpooling_layer()
helper.add_flatten_layer()
helper.add_fullyconn_layer(10)
helper.add_output_layer()
dq_network = DeepQLearning(action_wrapper=action_wrapper,
state_builder=state_builder,
build_model=helper.get_model_layout(),
gamma=0.99,
learning_rate=0.01,
epsilon_decay=0.9999,
epsilon_min=0.005,
memory_maxlen=50000,
min_memory_size=2000,
lib="keras")
agent = GenericAgent(dq_network, PureReward())
trainer = Trainer(env,
agent,
save_path='urnai/models/saved',
file_name="breakout-v0_keras",
save_every=100,
enable_save=True,
relative_path=True,
max_training_episodes=50,
max_steps_training=800,
max_test_episodes=5,
max_steps_testing=800)
return trainer
def declare_trainer():
env = SC2Env(map_name="Simple64",
render=False,
step_mul=16,
player_race="terran",
enemy_race="random",
difficulty="very_easy")
action_wrapper = SimpleTerranWrapper()
state_builder = Simple64GridState(grid_size=4)
helper = ModelBuilder()
helper.add_input_layer(nodes=50)
helper.add_fullyconn_layer(50)
helper.add_output_layer()
dq_network = DDQNKeras(action_wrapper=action_wrapper,
state_builder=state_builder,
build_model=helper.get_model_layout(),
per_episode_epsilon_decay=False,
gamma=0.99,
learning_rate=0.001,
epsilon_decay=0.99999,
epsilon_min=0.005,
memory_maxlen=100000,
min_memory_size=2000)
agent = SC2Agent(dq_network, KilledUnitsReward())
trainer = Trainer(env,
agent,
save_path='urnai/models/saved',
file_name="terran_ddqn_vs_random_v_easy",
save_every=50,
enable_save=True,
relative_path=True,
max_training_episodes=3000,
max_steps_training=1200,
max_test_episodes=100,
max_steps_testing=1200)
return trainer
def declare_trainer():
env = GymEnv(id="CartPole-v1")
action_wrapper = env.get_action_wrapper()
state_builder = GymState(env.env_instance.observation_space.shape[0])
helper = ModelBuilder()
helper.add_input_layer(nodes=50)
helper.add_fullyconn_layer(50)
helper.add_output_layer()
dq_network = DDQNKeras(action_wrapper=action_wrapper,
state_builder=state_builder,
build_model=helper.get_model_layout(),
use_memory=False,
gamma=0.99,
learning_rate=0.001,
epsilon_decay=0.9997,
epsilon_min=0.01,
memory_maxlen=50000,
min_memory_size=1000)
# dq_network = PGKeras(action_wrapper, state_builder, learning_rate=0.001, gamma=0.99, build_model=helper.get_model_layout())
agent = GenericAgent(dq_network, PureReward())
# Cartpole-v1 is solved when avg. reward over 100 episodes is greater than or equal to 475
#test_params = TestParams(num_matches=100, steps_per_test=100, max_steps=500, reward_threshold=500)
trainer = Trainer(env,
agent,
save_path='urnai/models/saved',
file_name="cartpole_v1_ddqn_50x50_test",
save_every=100,
enable_save=True,
relative_path=True,
max_training_episodes=1000,
max_steps_training=500,
max_test_episodes=100,
max_steps_testing=500)
return trainer
def declare_trainer():
env = GymEnv(id="CartPole-v0")
action_wrapper = env.get_action_wrapper()
state_builder = GymState(env.env_instance.observation_space.shape[0])
helper = ModelBuilder()
helper.add_input_layer()
helper.add_fullyconn_layer(25)
helper.add_output_layer()
dq_network = DDQNKeras(action_wrapper=action_wrapper,
state_builder=state_builder,
build_model=helper.get_model_layout(),
use_memory=False,
gamma=0.99,
learning_rate=0.001,
epsilon_decay=0.9997,
epsilon_min=0.01,
memory_maxlen=50000,
min_memory_size=1000)
agent = GenericAgent(dq_network, PureReward())
# Cartpole-v0 is solved when avg. reward over 100 episodes is greater than or equal to 195
trainer = Trainer(env,
agent,
save_path='urnai/models/saved',
file_name="cartpole-v0-unified-step-play3",
save_every=100,
enable_save=True,
relative_path=True,
max_training_episodes=1000,
max_steps_training=500,
max_test_episodes=100,
max_steps_testing=500)
return trainer
def declare_trainer():
env = GymEnv(id="FrozenLakeNotSlippery-v0")
action_wrapper = env.get_action_wrapper()
state_builder = FrozenLakeState()
training_date = str(datetime.now()).replace(" ", "_").replace(":",
"_").replace(
".", "_")
helper = ModelBuilder()
helper.add_input_layer()
helper.add_fullyconn_layer(256)
helper.add_fullyconn_layer(256)
helper.add_fullyconn_layer(256)
helper.add_fullyconn_layer(256)
helper.add_output_layer()
dq_network = DDQNKeras(action_wrapper=action_wrapper,
state_builder=state_builder,
learning_rate=0.005,
gamma=0.90,
use_memory=False,
per_episode_epsilon_decay=True,
build_model=helper.get_model_layout())
agent = GenericAgent(dq_network, FrozenlakeReward())
trainer = Trainer(env,
agent,
file_name=training_date + os.path.sep +
"frozenlake_test_ddqnKeras",
save_every=1000,
enable_save=True,
max_training_episodes=1000,
max_steps_training=500,
max_test_episodes=10,
max_steps_testing=500)
return trainer
def declare_trainer():
## Initializing our StarCraft 2 environment
env = SC2Env(map_name="Simple64", render=False, step_mul=16, player_race="terran", enemy_race="terran", difficulty="very_easy")
#action_wrapper = MOspatialTerranWrapper(10, 10, env.env_instance._interface_formats[0]._raw_resolution)
action_wrapper = SimpleMOTerranWrapper(10, 10, env.env_instance._interface_formats[0]._raw_resolution)
#state_builder = Simple64GridState_SimpleTerran(grid_size=7) # This state_builder with grid_size=7 will end upt with a total size of 110 ( (7*7)*2 + 12 )
state_builder = TVTUnitStackingEnemyGridState(grid_size=4) # total size of 92
helper = ModelBuilder()
helper.add_input_layer()
helper.add_fullyconn_layer(nodes=82)
helper.add_output_layer()
dq_network = DDQNKerasMO(action_wrapper=action_wrapper, state_builder=state_builder, build_model=helper.get_model_layout(), per_episode_epsilon_decay=False,
learning_rate=0.001, epsilon_decay=0.99999, epsilon_min=0.005, memory_maxlen=100000, min_memory_size=2000, batch_size=32)
# Terran agent
agent = SC2Agent(dq_network, KilledUnitsReward())
# trainer = Trainer(env, agent, save_path='/home/lpdcalves/', file_name="terran_ddqn_mo_v_easy",
# save_every=200, enable_save=True, relative_path=False, reset_epsilon=False,
# max_training_episodes=3000, max_steps_training=1200,
# max_test_episodes=100, max_steps_testing=1200, rolling_avg_window_size=50)
trainer = Trainer(env, agent, save_path='urnai/models/saved', file_name="terran_ddqn_mostackgridstate_atkgrouping4",
save_every=20, enable_save=True, relative_path=True, reset_epsilon=False,
max_training_episodes=10, max_steps_training=1200,
max_test_episodes=2, max_steps_testing=1200, rolling_avg_window_size=5)
return trainer