def __init__(self, config):
Base_Agent.__init__(self, config)
self.policy_output_size = self.calculate_policy_output_size()
self.policy_new = self.create_NN(input_dim=self.state_size,
output_dim=self.policy_output_size)
model_path = self.config.model_path if self.config.model_path else 'Models'
self.policy_new_path = os.path.join(
model_path, "{}_policy_new.pt".format(self.agent_name))
if self.config.load_model: self.locally_load_policy()
self.policy_old = self.create_NN(input_dim=self.state_size,
output_dim=self.policy_output_size)
self.policy_old.load_state_dict(
copy.deepcopy(self.policy_new.state_dict()))
self.policy_new_optimizer = optim.Adam(
self.policy_new.parameters(),
lr=self.hyperparameters["learning_rate"],
eps=1e-4)
self.episode_number = 0
self.many_episode_states = []
self.many_episode_actions = []
self.many_episode_rewards = []
self.experience_generator = Parallel_Experience_Generator(
self.environment, self.policy_new, self.config.seed,
self.hyperparameters, self.action_size)
self.exploration_strategy = Epsilon_Greedy_Exploration(self.config)
def __init__(self, config):
Base_Agent.__init__(self, config)
base_config.no_render_mode = False ## must be render mode
self.q_network_local = q_network_2_EYE(n_action=self.get_action_size())
self.q_network_target = q_network_2_EYE(
n_action=self.get_action_size())
self.q_network_optimizer = optim.SGD(
self.q_network_local.parameters(),
lr=self.hyperparameters["learning_rate"],
weight_decay=5e-4)
self.memory = Replay_Buffer(self.hyperparameters["buffer_size"],
self.hyperparameters["batch_size"],
config.seed)
self.exploration_strategy = Epsilon_Greedy_Exploration(config)
if config.backbone_pretrain:
self.load_pretrain()
self.copy_model_over(from_model=self.q_network_local,
to_model=self.q_network_target)
self.q_network_local.to(self.q_network_local.device)
self.q_network_target.to(self.q_network_target.device)
def __init__(self, config):
Base_Agent.__init__(self, config)
self.memory = Replay_Buffer(self.hyperparameters["buffer_size"], self.hyperparameters["batch_size"], config.seed)
self.q_network_local = self.create_NN(input_dim=self.state_size, output_dim=self.action_size)
self.q_network_optimizer = optim.Adam(self.q_network_local.parameters(),
lr=self.hyperparameters["learning_rate"])
self.exploration_strategy = Epsilon_Greedy_Exploration(config)
def __init__(self,
config,
global_action_id_to_primitive_actions,
action_length_reward_bonus,
end_of_episode_symbol="/"):
super().__init__(config)
self.end_of_episode_symbol = end_of_episode_symbol
self.global_action_id_to_primitive_actions = global_action_id_to_primitive_actions
self.memory = Replay_Buffer(self.hyperparameters["buffer_size"],
self.hyperparameters["batch_size"],
config.seed)
self.exploration_strategy = Epsilon_Greedy_Exploration(config)
self.oracle = self.create_oracle()
self.oracle_optimizer = optim.Adam(
self.oracle.parameters(), lr=self.hyperparameters["learning_rate"])
self.q_network_local = self.create_NN(input_dim=self.state_size + 1,
output_dim=self.action_size)
self.q_network_local.print_model_summary()
self.q_network_optimizer = optim.Adam(
self.q_network_local.parameters(),
lr=self.hyperparameters["learning_rate"])
self.q_network_target = self.create_NN(input_dim=self.state_size + 1,
output_dim=self.action_size)
Base_Agent.copy_model_over(from_model=self.q_network_local,
to_model=self.q_network_target)
self.action_length_reward_bonus = action_length_reward_bonus
self.abandon_ship = config.hyperparameters["abandon_ship"]
def __init__(self, config):
Base_Agent.__init__(self, config)
self.memory = Replay_Buffer(self.hyperparameters["buffer_size"],
self.hyperparameters["batch_size"],
config.seed)
self.q_network_local = Policy(self.state_size,
self.action_size).to("cuda")
self.q_network_optimizer = optim.Adam(
self.q_network_local.parameters(),
lr=self.hyperparameters["learning_rate"],
eps=1e-4)
self.exploration_strategy = Epsilon_Greedy_Exploration(config)
def __init__(self, config):
Base_Agent.__init__(self, config)
self.policy_output_size = self.calculate_policy_output_size()
self.policy_new = self.create_NN(input_dim=self.state_size, output_dim=self.policy_output_size)
self.policy_old = self.create_NN(input_dim=self.state_size, output_dim=self.policy_output_size)
self.policy_old.load_state_dict(copy.deepcopy(self.policy_new.state_dict()))
self.policy_new_optimizer = optim.Adam(self.policy_new.parameters(), lr=self.hyperparameters["learning_rate"])
self.episode_number = 0
self.many_episode_states = []
self.many_episode_actions = []
self.many_episode_rewards = []
self.experience_generator = Parallel_Experience_Generator(self.environment, self.policy_new, self.config.seed,
self.hyperparameters, self.action_size)
self.exploration_strategy = Epsilon_Greedy_Exploration(self.config)
def __init__(self, config, agent_name_=agent_name):
Base_Agent.__init__(self, config, agent_name=agent_name_)
self.memory = Replay_Buffer(self.hyperparameters["buffer_size"],
self.hyperparameters["batch_size"],
config.seed, self.device)
self.q_network_local = self.create_NN(
input_dim=self.state_size,
output_dim=self.action_size) # TODO: Change NN
self.q_network_optimizer = optim.Adam(
self.q_network_local.parameters(),
lr=self.hyperparameters["learning_rate"],
eps=1e-4)
self.exploration_strategy = Epsilon_Greedy_Exploration(config)
self.wandb_watch(self.q_network_local,
log_freq=self.config.wandb_model_log_freq)
def __init__(self, config):
Base_Agent.__init__(self, config)
model_path = self.config.model_path if self.config.model_path else 'Models'
self.memory = Replay_Buffer(self.hyperparameters["buffer_size"],
self.hyperparameters["batch_size"],
config.seed)
self.q_network_local = self.create_NN(input_dim=self.state_size,
output_dim=self.action_size)
self.q_network_local_path = os.path.join(
model_path, "{}_q_network_local.pt".format(self.agent_name))
if self.config.load_model: self.locally_load_policy()
self.q_network_optimizer = optim.Adam(
self.q_network_local.parameters(),
lr=self.hyperparameters["learning_rate"],
eps=1e-4)
self.exploration_strategy = Epsilon_Greedy_Exploration(config)
def __init__(self, config, agent_name_=agent_name):
Base_Agent.__init__(self, config, agent_name=agent_name_)
self.memory = Replay_Buffer(self.hyperparameters["buffer_size"],
self.hyperparameters["batch_size"],
config.seed, self.device)
# If model is not provided, create one. TODO Add this mechanism to all agents.
if not "model" in self.hyperparameters or self.hyperparameters[
"model"] is None:
self.q_network_local = self.create_NN(input_dim=self.state_size,
output_dim=self.action_size)
else:
self.q_network_local = self.hyperparameters["model"]
self.wandb_watch(self.q_network_local,
log_freq=self.config.wandb_model_log_freq)
self.q_network_optimizer = optim.Adam(
self.q_network_local.parameters(),
lr=self.hyperparameters["learning_rate"],
eps=1e-4)
self.exploration_strategy = Epsilon_Greedy_Exploration(config)
def __init__(self, config):
Base_Agent.__init__(self, config)
self.agent_dic = self.create_agent_dic()
self.exploration_strategy = Epsilon_Greedy_Exploration(config)
