def initialAgents(self):
# If no initial pelican agent is given, we train one from fresh
if len(self.initial_pelicans) == 0:
# Train initial pelican vs default panther
self.pelican_model = helper.make_new_model(self.model_type,
self.policy,
self.pelican_env,
n_steps=self.training_steps)
logger.info('Training initial pelican')
pelican_agent_filepath = self.train_agent(self.pelicans_tmp_exp_path,
self.pelican_model,
self.pelican_env)
else:
logger.info('Initial set of %d pelicans found' % (len(self.initial_pelicans)))
pelican_agent_filepath = self.initial_pelicans
# If no initial panther agent is given, we train one from fresh
if len(self.initial_panthers) == 0:
# Train initial panther agent vs default pelican
self.panther_model = helper.make_new_model(self.model_type,
self.policy,
self.panther_env,
n_steps=self.training_steps)
logger.info('Training initial panther')
panther_agent_filepath = self.train_agent(self.panthers_tmp_exp_path,
self.panther_model,
self.panther_env)
else:
logger.info('Initial set of %d panthers found' % (len(self.initial_panthers)))
panther_agent_filepath = self.initial_panthers
return panther_agent_filepath, pelican_agent_filepath
def bootstrap(self, model_paths, env, mixture):
if np.random.rand(1) < self.retraining_prob:
path = np.random.choice(model_paths, 1, p = mixture)[0]
path = glob.glob(path + "/*.zip")[0]
return helper.loadAgent(path, self.model_type)
else:
return helper.make_new_model(self.model_type,
self.policy,
env,
n_steps=self.training_steps)
def evaluate_algorithms(exp_name, base_path, tb_enabled, algorithms,
victory_threshold, victory_trials, max_seconds,
testing_interval, use_non_image):
basicdate = str(datetime.now().strftime("%Y%m%d_%H%M%S"))
exp_name = "{}_{}".format(exp_name, basicdate)
exp_path = os.path.join(base_path, exp_name)
logger.info("Storing results in {}".format(exp_path))
writer = None
if tb_enabled:
writer = SummaryWriter(exp_path)
for algo in algorithms:
tb_log_name = "{}_non_image".format(algo) if use_non_image else algo
logger.info("Evaluating algorithm: {}; non-image: {}".format(
algo, use_non_image))
if use_non_image:
image_based = False
env = plark_env_non_image_state.PlarkEnvNonImageState(
driving_agent='pelican',
config_file_path=
'/Components/plark-game/plark_game/game_config/10x10/panther_easy.json'
)
policy = "MlpPolicy" # CnnPolicy doesn't work with MultiDiscrete observation space
else:
image_based = True
env = plark_env.PlarkEnv(
driving_agent='pelican',
config_file_path=
'/Components/plark-game/plark_game/game_config/10x10/panther_easy.json'
)
policy = "CnnPolicy"
model = helper.make_new_model(algo, policy, env)
helper.train_until(model,
env,
victory_threshold,
victory_trials,
max_seconds,
testing_interval,
tb_writer=writer,
tb_log_name=tb_log_name)
helper.save_model_with_env_settings(exp_path, model, algo, env,
image_based, basicdate)
writer.close()
def compare_envs(exp_name, base_path, tb_enabled, victory_threshold,
victory_trials, max_seconds, testing_interval,
num_parallel_envs, non_image):
basicdate = str(datetime.now().strftime("%Y%m%d_%H%M%S"))
exp_name = "{}_{}".format(exp_name, basicdate)
exp_path = os.path.join(base_path, exp_name)
logger.info("Storing results in {}".format(exp_path))
writer = None
if tb_enabled:
writer = SummaryWriter(exp_path)
for parallel in [False, True]:
algo = "PPO2"
policy = "MlpPolicy" if non_image else "CnnPolicy"
tb_log_name = "{}_parallel".format(algo) if parallel else algo
logger.info("Evaluating {}; parallel: {}".format(algo, parallel))
if parallel:
logger.info("Evaluating using {} parallel environments".format(
num_parallel_envs))
env_fn = createNonImageEnv if non_image else createImageEnv
env = SubprocVecEnv([env_fn for _ in range(num_parallel_envs)])
else:
env = createNonImageEnv() if non_image else createImageEnv()
model = helper.make_new_model(algo, policy, env)
helper.train_until(model,
env,
victory_threshold,
victory_trials,
max_seconds,
testing_interval,
tb_writer=writer,
tb_log_name=tb_log_name)
helper.save_model_with_env_settings(exp_path, model, algo, env,
basicdate)
writer.close()
# +
pelican_env = PlarkEnvSparse(driving_agent='pelican',
config_file_path=config_file_path,
image_based=False,
random_panther_start_position=True,
max_illegal_moves_per_turn=1)
panther_env = PlarkEnvSparse(driving_agent='panther',
config_file_path=config_file_path,
image_based=False,
random_panther_start_position=True,
max_illegal_moves_per_turn=1)
# -
panthers = [
helper.make_new_model(model_type, policy_panther, panther_env)
for i in range(population_size)
]
pelicans = [
helper.make_new_model(model_type, policy_pelican, pelican_env)
for i in range(population_size)
]
# for iteration in tqdm(range(iterations)):
for iteration in range(iterations):
print("Iteration: " + str(iteration))
for panther in panthers:
for pelican in pelicans:
panther_env.set_pelican(pelican)
pelican_env.set_panther(panther)
pelican.learn(steps)
def run_self_play(exp_name,
exp_path,
basicdate,
pelican_testing_interval=100,
pelican_max_initial_learning_steps=10000,
panther_testing_interval=100,
panther_max_initial_learning_steps=10000,
self_play_testing_interval=100,
self_play_max_learning_steps_per_agent=10000,
self_play_iterations=10000,
model_type='PPO2',
log_to_tb=False,
image_based=True,
num_parallel_envs=1):
pelican_training_steps = 0
panther_training_steps = 0
pelican_model_type = model_type
panther_model_type = model_type
if log_to_tb:
writer = SummaryWriter(exp_path)
pelican_tb_log_name = 'pelican'
panther_tb_log_name = 'panther'
else:
writer = None
pelican_tb_log_name = None
panther_tb_log_name = None
policy = 'CnnPolicy'
if image_based is False:
policy = 'MlpPolicy'
parallel = False
if model_type.lower() == 'ppo2':
parallel = True
#Train initial pelican vs rule based panther
if parallel:
pelican_env = SubprocVecEnv([
lambda: PlarkEnv(
driving_agent='pelican',
config_file_path=
'/Components/plark-game/plark_game/game_config/10x10/pelican_easy.json',
image_based=image_based,
random_panther_start_position=True,
max_illegal_moves_per_turn=3) for _ in range(num_parallel_envs)
])
else:
pelican_env = PlarkEnv(
driving_agent='pelican',
config_file_path=
'/Components/plark-game/plark_game/game_config/10x10/pelican_easy.json',
image_based=image_based,
random_panther_start_position=True,
max_illegal_moves_per_turn=3)
pelican_model = helper.make_new_model(model_type, policy, pelican_env)
logger.info('Training initial pelican')
pelican_agent_filepath, steps = train_agent(
exp_path, pelican_model, pelican_env, pelican_testing_interval,
pelican_max_initial_learning_steps, pelican_model_type, basicdate,
writer, pelican_tb_log_name)
pelican_training_steps = pelican_training_steps + steps
# Train initial panther agent vs initial pelican agent
if parallel:
panther_env = SubprocVecEnv([
lambda: PlarkEnv(
driving_agent='panther',
pelican_agent_filepath=pelican_agent_filepath,
config_file_path=
'/Components/plark-game/plark_game/game_config/10x10/balanced.json',
image_based=image_based,
random_panther_start_position=True,
max_illegal_moves_per_turn=3) for _ in range(num_parallel_envs)
])
else:
panther_env = PlarkEnv(
driving_agent='panther',
pelican_agent_filepath=pelican_agent_filepath,
config_file_path=
'/Components/plark-game/plark_game/game_config/10x10/balanced.json',
image_based=image_based,
random_panther_start_position=True,
max_illegal_moves_per_turn=3)
panther_model = helper.make_new_model(model_type, policy, panther_env)
logger.info('Training initial panther')
panther_agent_filepath, steps = train_agent(
exp_path, panther_model, panther_env, panther_testing_interval,
panther_max_initial_learning_steps, panther_model_type, basicdate,
writer, panther_tb_log_name)
panther_training_steps = panther_training_steps + steps
# Train agent vs agent
logger.info('Self play')
for i in range(self_play_iterations):
logger.info('Self play iteration ' + str(i) + ' of ' +
str(self_play_iterations))
logger.info('Training pelican')
if parallel:
pelican_env = SubprocVecEnv([
lambda: PlarkEnvSparse(
driving_agent='pelican',
panther_agent_filepath=panther_agent_filepath,
config_file_path=
'/Components/plark-game/plark_game/game_config/10x10/balanced.json',
image_based=image_based,
random_panther_start_position=True,
max_illegal_moves_per_turn=3)
for _ in range(num_parallel_envs)
])
else:
pelican_env = PlarkEnvSparse(
driving_agent='pelican',
panther_agent_filepath=panther_agent_filepath,
config_file_path=
'/Components/plark-game/plark_game/game_config/10x10/balanced.json',
image_based=image_based,
random_panther_start_position=True,
max_illegal_moves_per_turn=3)
pelican_agent_filepath, steps = train_agent(
exp_path,
pelican_model,
pelican_env,
self_play_testing_interval,
self_play_max_learning_steps_per_agent,
pelican_model_type,
basicdate,
writer,
pelican_tb_log_name,
previous_steps=pelican_training_steps)
pelican_training_steps = pelican_training_steps + steps
logger.info('Training panther')
if parallel:
panther_env = SubprocVecEnv([
lambda: PlarkEnvSparse(
driving_agent='panther',
pelican_agent_filepath=pelican_agent_filepath,
config_file_path=
'/Components/plark-game/plark_game/game_config/10x10/balanced.json',
image_based=image_based,
random_panther_start_position=True,
max_illegal_moves_per_turn=3)
for _ in range(num_parallel_envs)
])
else:
panther_env = PlarkEnvSparse(
driving_agent='panther',
pelican_agent_filepath=pelican_agent_filepath,
config_file_path=
'/Components/plark-game/plark_game/game_config/10x10/balanced.json',
image_based=image_based,
random_panther_start_position=True,
max_illegal_moves_per_turn=3)
panther_agent_filepath, steps = train_agent(
exp_path,
panther_model,
panther_env,
self_play_testing_interval,
self_play_max_learning_steps_per_agent,
panther_model_type,
basicdate,
writer,
panther_tb_log_name,
previous_steps=panther_training_steps)
panther_training_steps = panther_training_steps + steps
logger.info('Training pelican total steps:' + str(pelican_training_steps))
logger.info('Training panther total steps:' + str(panther_training_steps))
# Make video
video_path = os.path.join(exp_path, 'test_self_play.mp4')
basewidth, hsize = helper.make_video(pelican_model, pelican_env,
video_path)
return video_path, basewidth, hsize
# +
pelican_env = envops(
PlarkEnvSparse(driving_agent='pelican',
config_file_path=config_file_path,
image_based=False,
random_panther_start_position=True,
max_illegal_moves_per_turn=1), log_dir_base + '/pelican/')
panther_env = envops(
PlarkEnvSparse(driving_agent='panther',
config_file_path=config_file_path,
image_based=False,
random_panther_start_position=True,
max_illegal_moves_per_turn=1), log_dir_base + '/panther/')
pelican = helper.make_new_model(model_type, policy, pelican_env)
panther = helper.make_new_model(model_type, policy, panther_env)
panther_env.set_pelican(pelican)
pelican_env.set_panther(panther)
# -
for iteration in tqdm(range(iterations)):
pelican.learn(steps)
panther.learn(steps)
# Make video
video_path = os.path.join('./', 'test_self_play.mp4')
basewidth, hsize = helper.make_video(pelican, pelican_env, video_path)
video = io.open(video_path, 'r+b').read()
encoded = base64.b64encode(video)
