def log_training_start_information(self):
text = ("\n Agent: {}\n".format(self.agent_name) +
" ActionWrapper: {}\n".format(self.action_wrapper_name) +
" StateBuilder: {}\n".format(self.state_builder_name) +
" RewardBuilder: {}\n".format(self.reward_builder_name) +
" Environment: {}\n".format(self.env_name) +
" Model: {}\n".format(self.model_name))
if (hasattr(self.model, "lib")):
if self.model.neural_net_class != None:
if self.model.lib == constants.Libraries.KERAS:
stringlist = []
self.model.dnn.model.summary(
print_fn=lambda x: stringlist.append(x))
short_model_summary = "\n".join(stringlist)
text += " " + short_model_summary
if self.model.lib == constants.Libraries.PYTORCH:
text += " " + self.model.dnn.model
else:
for idx, (layer) in enumerate(self.model.build_model):
text += " Layer {}: {}\n".format(
idx, self.model.build_model[idx])
else:
for idx, (layer) in enumerate(self.model.build_model):
text += " Layer {}: {}\n".format(
idx, self.model.build_model[idx])
self.training_report += text
rp.report(text)
def save(self, savepath):
'''
This method saves pickle objects
and extra stuff needed
'''
rp.report("Saving {} object...".format(self.__class__.__name__), verbosity_lvl=1)
self.save_pickle(savepath)
self.save_extra(savepath)
def ask_for_continue(self):
if self.version != self.__curr_version:
answer = ""
while answer.lower() != "y" and answer.lower() != "n":
answer = rp.input("The loaded training version is {} and the current version is {}. This difference can cause some kind of error while proceeding to the training, do you wish to continue? [y/n]".format(self.version, self.__curr_version), "n")
if answer.lower() == "n":
rp.report("The training was stopped.")
exit()
def get_sc2_reward(self, obs):
build_supply_depot = BuildUnitsGeneralizedRewardBuilder.ACTION_BUILD_SUPPLY_DEPOT
build_barrack = BuildUnitsGeneralizedRewardBuilder.ACTION_BUILD_BARRACK
build_marine = BuildUnitsGeneralizedRewardBuilder.ACTION_BUILD_MARINE
do_nothing = BuildUnitsGeneralizedRewardBuilder.ACTION_DO_NOTHING
current = self.get_sc2_number_of_supply_depot(obs)
prev = self.get_sc2_number_of_supply_depot(self.previous_state)
supply_depot_amount_diff = (current - prev)
current = self.get_sc2_number_of_barracks(obs)
prev = self.get_sc2_number_of_barracks(self.previous_state)
barracks_amount_diff = (current - prev)
current = self.get_sc2_number_of_marines(obs)
prev = self.get_sc2_number_of_marines(self.previous_state)
marines_amount_diff = (current - prev)
negative_rwd = 0
chosen_action = BuildUnitsGeneralizedRewardBuilder.LAST_CHOSEN_ACTION
if chosen_action > -1:
supply_depot_amount = self.get_sc2_number_of_supply_depot(obs)
barracks_amount = self.get_sc2_number_of_barracks(obs)
minerals = obs.player.minerals
if chosen_action == build_supply_depot:
if supply_depot_amount > 7 or minerals < 100:
negative_rwd = -10
elif chosen_action == build_barrack:
if supply_depot_amount <= 0 or minerals < 150:
negative_rwd = -10
elif chosen_action == build_marine:
if barracks_amount <= 0 or minerals < 50:
negative_rwd = -10
#elif chosen_action == do_nothing:
# negative_rwd = -1
#rwd = negative_rwd + rwdB + rwdC
rp.report('''
Calculated reward is: {},
composed of:
supply_depot_amount: {},
barracks_amount: {},
marines_amount: {},
negative_rdw: {}
'''.format(
negative_rwd + supply_depot_amount_diff +
barracks_amount_diff * 10 + marines_amount_diff * 100,
supply_depot_amount_diff, barracks_amount_diff * 10,
marines_amount_diff * 100, negative_rwd),
verbosity_lvl=1)
if supply_depot_amount_diff < 0 or barracks_amount_diff < 0 or marines_amount_diff < 0:
return 0
else:
rwd = negative_rwd + supply_depot_amount_diff + barracks_amount_diff * 10 + marines_amount_diff * 100
return rwd
def load_pickle(self, persist_path):
'''
This method loads a list instance
saved by pickle.
'''
#Check if pickle file exists
pickle_path = self.get_full_persistance_pickle_path(persist_path)
exists_pickle = os.path.isfile(pickle_path)
#If yes, load it
if exists_pickle:
if os.path.getsize(pickle_path) > 0:
with open(pickle_path, "rb") as pickle_in:
pickle_dict = pickle.load(pickle_in)
self.restore_pickleable_attributes(pickle_dict)
rp.report("**************************************** \n Pickle for " + self.get_default_save_stamp() + " loaded. \n****************************************", 1)
def log_train_stats(self):
if self.ep_count > 0:
text = ("\n" + "Current Reward Avg.: {}".format(
sum(self.ep_rewards) / self.ep_count) +
" Win rate: {:10.3f}%".format(
(sum(self.ep_victories) / self.ep_count) * 100) +
" Avg number of steps: {}".format(
sum(self.ep_avg_steps) / self.ep_count) +
" Training Duration (seconds): {}".format(
round(time() - self.training_start, 2)) + "\n")
self.training_report += text
rp.report(text)
else:
rp.report("There are no recorded episodes!")
def step(self, action):
if (self.game == GeneralizedBuildUnitsScenario.GAME_DEEP_RTS):
BuildUnitsGeneralizedRewardBuilder.LAST_CHOSEN_ACTION = action
if self.steps == 0:
self.setup_map()
self.spawn_army()
elif self.steps == 1:
self.collect_gold()
if rp.VERBOSITY_LEVEL > 0:
str_ = ''' DRTS Episode Status:
Number of gold = {},
Number of barracks = {},
Number of farms = {},
Number of soldiers = {}'''.format(
self.env.players[0].gold,
self.get_drts_unit_type_count(0, self.env.constants.Unit.Barracks),
self.get_drts_unit_type_count(0, self.env.constants.Unit.Farm),
self.get_drts_unit_type_count(0, self.env.constants.Unit.Footman),
)
rp.report(str_, verbosity_lvl=1)
state, reward, done = None, None, None
if action == GeneralizedBuildUnitsScenario.ACTION_DRTS_DO_NOTHING:
no_action = 15
state, reward, done = self.env.step(no_action)
elif action == GeneralizedBuildUnitsScenario.ACTION_DRTS_BUILD_FARM:
self.build_farm()
no_action = 15
state, reward, done = self.env.step(no_action)
elif action == GeneralizedBuildUnitsScenario.ACTION_DRTS_BUILD_BARRACK:
self.build_barrack()
no_action = 15
state, reward, done = self.env.step(no_action)
elif action == GeneralizedBuildUnitsScenario.ACTION_DRTS_BUILD_FOOTMAN:
self.build_footman()
no_action = 15
state, reward, done = self.env.step(no_action)
else:
state, reward, done = self.env.step(action)
self.steps += 1
return state, reward, done
elif (self.game == GeneralizedBuildUnitsScenario.GAME_STARCRAFT_II):
self.steps += 1
return self.env.step(action)
def log_ep_stats(self):
if self.ep_count > 0:
agent_info = dict.fromkeys(self.agent_info)
for key in agent_info:
agent_info[key] = self.agent_info[key][-1]
rp.report(
"Episode: {}/{} | Outcome: {} | Episode Avg. Reward: {:10.6f} | Episode Reward: {:10.6f} | Episode Steps: {:10.6f} | Best Reward was {} on episode: {} | Episode Duration (seconds): {} | Episode SPS: {} | SPS AVG: {} | Agent info: {}"
.format(self.ep_count, self.ep_total, self.ep_victories[-1],
self.ep_avg_rewards[-1], self.ep_rewards[-1],
self.ep_steps_count[-1], self.best_reward,
self.best_reward_episode,
self.episode_duration_list[-1],
self.episode_sps_list[-1], self.avg_sps_list[-1],
agent_info))
else:
rp.report("There are no recorded episodes!")
def test_agent(self):
#backup attributes
max_test_episodes_backup = self.max_test_episodes
curr_playing_episodes_backup = self.curr_playing_episodes
logger_backup = self.logger
#full_save_play_path_backup = self.full_save_play_path
enable_save_backup = self.enable_save
#set attributes to test agent
self.enable_save = False
#self.full_save_play_path = self.full_save_path + os.path.sep + "inside_training_play_files" + os.path.sep + "test_at_training_episode_{}".format(self.curr_training_episodes)
#self.make_persistance_dirs(self.log_actions)
self.max_test_episodes = self.reward_test_number_of_episodes
self.curr_playing_episodes = 0
rp.report("> Starting to check current agent performance.")
#make the agent play
self.play()
rp.report("> Finished checking current agent performance.")
#get_reward_avg
rwd_avg = self.logger.ep_avg_rewards[-1]
#save this logger for later saving
#this is needed to get some more detailed
#info on tests
logger_dict = {}
logger_dict["logger"] = self.logger
logger_dict["saved"] = False
self.inside_training_test_loggers.append(logger_dict)
#restore backup
self.max_test_episodes = max_test_episodes_backup
self.curr_playing_episodes = curr_playing_episodes_backup
self.logger = logger_backup
#self.full_save_play_path = full_save_play_path_backup
self.enable_save = enable_save_backup
#register reward avg:
self.logger.inside_training_test_avg_rwds.append(rwd_avg)
def setup(self,
env,
agent,
max_training_episodes,
max_test_episodes,
max_steps_training,
max_steps_testing,
save_path=os.path.expanduser("~") + os.path.sep +
"urnai_saved_traingings",
file_name=str(datetime.now()).replace(" ", "_").replace(
":", "_").replace(".", "_"),
enable_save=True,
save_every=10,
relative_path=False,
debug_level=0,
reset_epsilon=False,
tensorboard_logging=False,
log_actions=True,
episode_batch_avg_calculation=10,
do_reward_test=False,
reward_test_number_of_episodes=10,
rolling_avg_window_size=20):
self.versioner = Versioner()
self.env = env
self.agent = agent
self.save_path = save_path
self.file_name = file_name
self.enable_save = enable_save
self.save_every = save_every
self.relative_path = relative_path
self.reset_epsilon = reset_epsilon
self.max_training_episodes = max_training_episodes
self.max_test_episodes = max_test_episodes
self.max_steps_training = max_steps_training
self.max_steps_testing = max_steps_testing
self.curr_training_episodes = -1
self.curr_playing_episodes = -1
rp.VERBOSITY_LEVEL = debug_level
self.tensorboard_logging = tensorboard_logging
self.log_actions = log_actions
self.episode_batch_avg_calculation = episode_batch_avg_calculation
self.do_reward_test = do_reward_test
self.reward_test_number_of_episodes = reward_test_number_of_episodes
self.rolling_avg_window_size = rolling_avg_window_size
self.inside_training_test_loggers = []
self.logger = Logger(
0,
self.agent.__class__.__name__,
self.agent.model.__class__.__name__,
self.agent.model,
self.agent.action_wrapper.__class__.__name__,
self.agent.action_wrapper.get_action_space_dim(),
self.agent.action_wrapper.get_named_actions(),
self.agent.state_builder.__class__.__name__,
self.agent.reward_builder.__class__.__name__,
self.env.__class__.__name__,
log_actions=self.log_actions,
episode_batch_avg_calculation=self.episode_batch_avg_calculation,
rolling_avg_window_size=self.rolling_avg_window_size)
# Adding epsilon, learning rate and gamma factors to our pickle black list,
# so that they are not loaded when loading the model's weights.
# Making it so that the current training session acts as a brand new training session
# (except for the fact that the model's weights may already be somewhat optimized from previous trainings)
if self.reset_epsilon:
self.agent.model.pickle_black_list.append("epsilon_greedy")
self.agent.model.pickle_black_list.append("epsilon_decay_rate")
self.agent.model.pickle_black_list.append("epsilon_min")
self.agent.model.pickle_black_list.append("gamma")
self.agent.model.pickle_black_list.append("learning_rate")
self.agent.model.pickle_black_list.append("learning_rate_min")
self.agent.model.pickle_black_list.append("learning_rate_decay")
self.agent.model.pickle_black_list.append(
"learning_rate_decay_ep_cutoff")
currentdir = os.path.dirname(
os.path.abspath(inspect.getfile(inspect.currentframe())))
parentdir = os.path.dirname(currentdir)
parentdir = os.path.dirname(parentdir)
if (relative_path):
self.full_save_path = parentdir + os.path.sep + self.save_path + os.path.sep + self.file_name
else:
self.full_save_path = self.save_path + os.path.sep + self.file_name
self.full_save_play_path = self.full_save_path + os.path.sep + "play_files"
if self.enable_save and os.path.exists(self.full_save_path):
rp.report("WARNING! Loading training from " + self.full_save_path +
" with SAVING ENABLED.")
self.load(self.full_save_path)
self.versioner.ask_for_continue()
self.make_persistance_dirs(self.log_actions)
elif self.enable_save:
rp.report("WARNING! Starting new training on " +
self.full_save_path + " with SAVING ENABLED.")
self.make_persistance_dirs(self.log_actions)
else:
rp.report(
"WARNING! Starting new training WITHOUT SAVING PROGRESS.")
if (self.tensorboard_logging):
logdir = self.full_save_path + "/tf_logs"
self.agent.model.tensorboard_callback = [
tf.keras.callbacks.TensorBoard(log_dir=logdir)
]
def training_loop(self, is_testing, reward_from_agent=True):
start_time = time.time()
#current_episodes = 0
if is_testing:
rp.report("\n\n> Playing")
max_episodes = self.max_test_episodes
max_steps = self.max_steps_testing
current_episodes = self.curr_playing_episodes
else:
rp.report("> Training")
max_episodes = self.max_training_episodes
max_steps = self.max_steps_training
current_episodes = self.curr_training_episodes
if self.logger.ep_count == 0 or is_testing:
self.logger = Logger(
max_episodes,
self.agent.__class__.__name__,
self.agent.model.__class__.__name__,
self.agent.model,
self.agent.action_wrapper.__class__.__name__,
self.agent.action_wrapper.get_action_space_dim(),
self.agent.action_wrapper.get_named_actions(),
self.agent.state_builder.__class__.__name__,
self.agent.reward_builder.__class__.__name__,
self.env.__class__.__name__,
log_actions=self.log_actions,
episode_batch_avg_calculation=self.
episode_batch_avg_calculation,
rolling_avg_window_size=self.rolling_avg_window_size)
while current_episodes < max_episodes:
current_episodes += 1
self.env.start()
if is_testing:
self.curr_playing_episodes = current_episodes
else:
self.curr_training_episodes = current_episodes
# Reset the environment
obs = self.env.reset()
step_reward = 0
done = False
# Passing the episode to the agent reset, so that it can be passed to model reset
# Allowing the model to track the episode number, and decide if it should diminish the
# Learning Rate, depending on the currently selected strategy.
self.agent.reset(current_episodes)
ep_reward = 0
victory = False
ep_actions = np.zeros(
self.agent.action_wrapper.get_action_space_dim())
self.logger.record_episode_start()
for step in range(max_steps):
# Choosing an action and passing it to our env.step() in order to act on our environment
action = self.agent.step(obs, done, is_testing)
# Take the action (a) and observe the outcome state (s') and reward (r)
obs, default_reward, done = self.env.step(action)
# Logic to test wheter this is the last step of this episode
is_last_step = step == max_steps - 1
done = done or is_last_step
# Checking whether or not to use the reward from the reward builder so we can pass that to the agent
if reward_from_agent:
step_reward = self.agent.get_reward(
obs, default_reward, done)
else:
step_reward = default_reward
# Making the agent learn
if not is_testing:
self.agent.learn(obs, step_reward, done)
# Adding our step reward to the total count of the episode's reward
ep_reward += step_reward
ep_actions[self.agent.previous_action] += 1
if done:
victory = default_reward == 1
agent_info = {
"Learning rate": self.agent.model.learning_rate,
"Gamma": self.agent.model.gamma,
"Epsilon": self.agent.model.epsilon_greedy,
}
self.logger.record_episode(ep_reward, victory, step + 1,
agent_info, ep_actions)
break
self.logger.log_ep_stats()
# check if user wants to pause training and test agent
# if self.do_reward_test and current_episodes % self.episode_batch_avg_calculation == 0 and current_episodes > 1:
if (
not is_testing
) and self.do_reward_test and current_episodes % self.episode_batch_avg_calculation == 0:
self.test_agent()
# if this is not a test (evaluation), saving is enabled and we are in a multiple
# of our save_every variable then we save the model and generate graphs
if (
not is_testing
) and self.enable_save and current_episodes > 0 and current_episodes % self.save_every == 0:
self.save(self.full_save_path)
# if we have done tests along the training save all loggers for further detailed analysis
if self.do_reward_test and len(
self.inside_training_test_loggers) > 0:
for idx in range(
len(self.logger.ep_avg_batch_rewards_episodes)):
logger_dict = self.inside_training_test_loggers[idx]
if not logger_dict["saved"]:
episode = self.logger.ep_avg_batch_rewards_episodes[
idx]
backup_full_save_path = self.full_save_path
self.full_save_path = self.full_save_path + os.path.sep + "inside_training_play_files" + os.path.sep + "test_at_training_episode_{}".format(
episode)
self.make_persistance_dirs(self.log_actions)
logger_dict["logger"].save(self.full_save_path)
logger_dict["saved"] = True
self.full_save_path = backup_full_save_path
end_time = time.time()
if is_testing:
rp.report("\n> Test duration: {} seconds".format(end_time -
start_time))
self.logger.log_train_stats()
else:
rp.report("\n> Training duration: {} seconds".format(end_time -
start_time))
self.logger.log_train_stats()
# Saving the model at the end of the training loop
if self.enable_save:
if is_testing:
self.logger.save(self.full_save_play_path)
rp.save(self.full_save_play_path)
else:
self.save(self.full_save_path)
# if we have done tests along the training save all loggers for further detailed analysis
if self.do_reward_test and len(
self.inside_training_test_loggers) > 0:
for idx in range(
len(self.logger.ep_avg_batch_rewards_episodes)):
logger_dict = self.inside_training_test_loggers[idx]
if not logger_dict["saved"]:
episode = self.logger.ep_avg_batch_rewards_episodes[
idx]
backup_full_save_path = self.full_save_path
self.full_save_path = self.full_save_path + os.path.sep + "inside_training_play_files" + os.path.sep + "test_at_training_episode_{}".format(
episode)
self.make_persistance_dirs(self.log_actions)
logger_dict["logger"].save(self.full_save_path)
logger_dict["saved"] = True
self.full_save_path = backup_full_save_path
def old_play(self, test_params=None, reward_from_agent=True):
rp.report("\n\n> Playing")
self.logger = Logger(
self.max_test_episodes,
self.agent.__class__.__name__,
self.agent.model.__class__.__name__,
self.agent.model,
self.agent.action_wrapper.__class__.__name__,
self.agent.action_wrapper.get_action_space_dim(),
self.agent.action_wrapper.get_named_actions(),
self.agent.state_builder.__class__.__name__,
self.agent.reward_builder.__class__.__name__,
self.env.__class__.__name__,
log_actions=self.log_actions,
episode_batch_avg_calculation=self.episode_batch_avg_calculation,
rolling_avg_window_size=self.rolling_avg_window_size)
while self.curr_playing_episodes < self.max_test_episodes:
self.curr_playing_episodes += 1
self.env.start()
# Reset the environment
obs = self.env.reset()
step_reward = 0
done = False
# Passing the episode to the agent reset, so that it can be passed to model reset
# Allowing the model to track the episode number, and decide if it should diminish the
# Learning Rate, depending on the currently selected strategy.
self.agent.reset(self.curr_playing_episodes)
ep_reward = 0
victory = False
ep_actions = np.zeros(
self.agent.action_wrapper.get_action_space_dim())
self.logger.record_episode_start()
for step in range(self.max_steps_testing):
action = self.agent.step(obs, done, is_testing=True)
# Take the action (a) and observe the outcome state(s') and reward (r)
obs, default_reward, done = self.env.step(action)
is_last_step = step == self.max_steps_testing - 1
done = done or is_last_step
if reward_from_agent:
step_reward = self.agent.get_reward(
obs, default_reward, done)
else:
step_reward = default_reward
ep_reward += step_reward
ep_actions[self.agent.previous_action] += 1
# If done: finish episode
if done:
victory = default_reward == 1
agent_info = {
"Learning rate": self.agent.model.learning_rate,
"Gamma": self.agent.model.gamma,
"Epsilon": self.agent.model.epsilon_greedy,
}
self.logger.record_episode(ep_reward, victory, step + 1,
agent_info, ep_actions)
break
self.logger.log_ep_stats()
if test_params != None:
test_params.logger.record_play_test(test_params.current_ep_count,
self.logger.ep_rewards,
self.logger.victories,
self.max_test_episodes)
else:
# Only logs train stats if this is not a test, to avoid cluttering the interface with info
self.logger.log_train_stats()
# We need to save playing status as well
if self.enable_save:
self.logger.save(self.full_save_play_path)
rp.save(self.full_save_play_path)
def old_train(self,
test_params: TestParams = None,
reward_from_agent=True):
start_time = time.time()
rp.report("> Training")
if self.logger.ep_count == 0:
self.logger = Logger(
self.max_training_episodes,
self.agent.__class__.__name__,
self.agent.model.__class__.__name__,
self.agent.model,
self.agent.action_wrapper.__class__.__name__,
self.agent.action_wrapper.get_action_space_dim(),
self.agent.action_wrapper.get_named_actions(),
self.agent.state_builder.__class__.__name__,
self.agent.reward_builder.__class__.__name__,
self.env.__class__.__name__,
log_actions=self.log_actions,
episode_batch_avg_calculation=self.
episode_batch_avg_calculation,
rolling_avg_window_size=self.rolling_avg_window_size)
if test_params != None:
test_params.logger = self.logger
while self.curr_training_episodes < self.max_training_episodes:
self.curr_training_episodes += 1
self.env.start()
# Reset the environment
obs = self.env.reset()
step_reward = 0
done = False
# Passing the episode to the agent reset, so that it can be passed to model reset
# Allowing the model to track the episode number, and decide if it should diminish the
# Learning Rate, depending on the currently selected strategy.
self.agent.reset(self.curr_training_episodes)
ep_reward = 0
victory = False
ep_actions = np.zeros(
self.agent.action_wrapper.get_action_space_dim())
self.logger.record_episode_start()
for step in range(self.max_steps_training):
# Choosing an action and passing it to our env.step() in order to act on our environment
action = self.agent.step(obs, done, is_testing=False)
obs, default_reward, done = self.env.step(action)
is_last_step = step == self.max_steps_training - 1
done = done or is_last_step
# Checking whether or not to use the reward from the reward builder so we can pass that to the agent
if reward_from_agent:
step_reward = self.agent.get_reward(
obs, default_reward, done)
else:
step_reward = default_reward
# Making the agent learn
self.agent.learn(obs, step_reward, done)
# Adding our step reward to the total count of the episode's reward
ep_reward += step_reward
ep_actions[self.agent.previous_action] += 1
if done:
victory = default_reward == 1
agent_info = {
"Learning rate": self.agent.model.learning_rate,
"Gamma": self.agent.model.gamma,
"Epsilon": self.agent.model.epsilon_greedy,
}
self.logger.record_episode(ep_reward, victory, step + 1,
agent_info, ep_actions)
break
self.logger.log_ep_stats()
#check if user wants to pause training and test agent
#if self.do_reward_test and self.curr_training_episodes % self.episode_batch_avg_calculation == 0 and self.curr_training_episodes > 1:
if self.do_reward_test and self.curr_training_episodes % self.episode_batch_avg_calculation == 0:
self.test_agent()
if self.enable_save and self.curr_training_episodes > 0 and self.curr_training_episodes % self.save_every == 0:
self.save(self.full_save_path)
#if we have done tests along the training
#save all loggers for further detailed analysis
#this was needed because the play() method
#was saving these loggers every test, slowing down
#training a lot. Putting this code here allows
#to save them once and optimize training time.
if self.do_reward_test and len(
self.inside_training_test_loggers) > 0:
for idx in range(
len(self.logger.ep_avg_batch_rewards_episodes)):
logger_dict = self.inside_training_test_loggers[idx]
if not logger_dict["saved"]:
episode = self.logger.ep_avg_batch_rewards_episodes[
idx]
backup_full_save_path = self.full_save_path
self.full_save_path = self.full_save_path + os.path.sep + "inside_training_play_files" + os.path.sep + "test_at_training_episode_{}".format(
episode)
self.make_persistance_dirs(self.log_actions)
logger_dict["logger"].save(self.full_save_path)
logger_dict["saved"] = True
self.full_save_path = backup_full_save_path
if test_params != None and self.curr_training_episodes % test_params.test_steps == 0 and episode != 0:
test_params.current_ep_count = self.curr_training_episodes
self.play(test_params.num_matches, test_params.max_steps,
test_params)
# Stops training if reward threshold was reached in play testing
if test_params.reward_threshold != None and test_params.reward_threshold <= test_params.logger.play_rewards_avg[
-1]:
rp.report("> Reward threshold was reached!")
rp.report("> Stopping training")
break
end_time = time.time()
rp.report("\n> Training duration: {} seconds".format(end_time -
start_time))
self.logger.log_train_stats()
self.logger.plot_train_stats()
# Saving the model when the training has ended
if self.enable_save:
self.save(self.full_save_path)
#if we have done tests along the training
#save all loggers for further detailed analysis
#this was needed because the play() method
#was saving these loggers every test, slowing down
#training a lot. Putting this code here allows
#to save them once and optimize training time.
if self.do_reward_test and len(
self.inside_training_test_loggers) > 0:
for idx in range(len(
self.logger.ep_avg_batch_rewards_episodes)):
logger_dict = self.inside_training_test_loggers[idx]
if not logger_dict["saved"]:
episode = self.logger.ep_avg_batch_rewards_episodes[
idx]
backup_full_save_path = self.full_save_path
self.full_save_path = self.full_save_path + os.path.sep + "inside_training_play_files" + os.path.sep + "test_at_training_episode_{}".format(
episode)
self.make_persistance_dirs(self.log_actions)
logger_dict["logger"].save(self.full_save_path)
logger_dict["saved"] = True
self.full_save_path = backup_full_save_path
def get_drts_reward(self, obs):
player = 0
footman = 5
farm = 6
barracks = 4
build_farm = RTSGeneralization.ACTION_DRTS_BUILD_FARM
build_barrack = RTSGeneralization.ACTION_DRTS_BUILD_BARRACK
build_footman = RTSGeneralization.ACTION_DRTS_BUILD_FOOTMAN
do_nothing = RTSGeneralization.ACTION_DRTS_DO_NOTHING
current = self.get_drts_number_of_specific_units(obs, player, farm)
prev = self.get_drts_number_of_specific_units(self.previous_state,
player, farm)
farm_amount_curr = (current - prev)
current = self.get_drts_number_of_specific_units(obs, player, barracks)
prev = self.get_drts_number_of_specific_units(self.previous_state,
player, barracks)
barracks_amount_curr = (current - prev)
current = self.get_drts_number_of_specific_units(obs, player, footman)
prev = self.get_drts_number_of_specific_units(self.previous_state,
player, footman)
footman_amount_curr = (current - prev)
negative_rwd = 0
chosen_action = BuildUnitsGeneralizedRewardBuilder.LAST_CHOSEN_ACTION
if chosen_action > -1:
farm_number = self.get_drts_number_of_specific_units(
obs, player, farm)
barracks_amount = self.get_drts_number_of_specific_units(
obs, player, barracks)
gold_amount = obs['players'][0].gold
if chosen_action == build_farm:
if farm_number > 7 or gold_amount < 500:
negative_rwd = -10
elif chosen_action == build_barrack:
if farm_number <= 0 or gold_amount < 700:
negative_rwd = -10
elif chosen_action == build_footman:
if barracks_amount <= 0 or gold_amount < 600:
negative_rwd = -10
#elif chosen_action == do_nothing:
# negative_rwd = -1
#rwd = negative_rwd + rwdB + rwdC
rp.report('''
Calculated reward is: {},
composed of:
farm_amount: {},
barracks_amount: {},
footman_amount: {},
negative_rdw: {}
'''.format(
negative_rwd + farm_amount_curr + barracks_amount_curr * 10 +
footman_amount_curr * 100, farm_amount_curr,
barracks_amount_curr * 10, footman_amount_curr * 100,
negative_rwd),
verbosity_lvl=1)
if farm_amount_curr < 0 or barracks_amount_curr < 0 or footman_amount_curr < 0:
return 0
else:
rwd = negative_rwd + farm_amount_curr + barracks_amount_curr * 10 + footman_amount_curr * 100
return rwd