def _init_alg(self):
"""
Initialize the algorithm based on what algorithm is specified.
"""
# init storage for actor and critic models
self.actors = []
self.actor_targets = []
self.critics = []
self.critic_targets = []
# create all models separately for each agent instance
for _ in range(self.num_instances):
actor = ActorNetwork(self.state_size, self.action_size)
target_actor = ActorNetwork(self.state_size, self.action_size)
target_actor = utils.copy_weights(actor)
critic = CriticNetwork(self.state_size, self.action_size)
target_critic = CriticNetwork(self.state_size, self.action_size)
target_critic = utils.copy_weights(critic)
self.actors.append(actor)
self.actor_targets.append(target_actor)
self.critics.append(critic)
self.critic_targets.append(target_critic)
# initialize the replay buffer
self.memory = ReplayBuffer(self.buffer_size,
self.batch_size,
seed=self.seed)
def _init_alg(self):
"""
Initialize the algorithm based on what algorithm is specified.
"""
# initialize the actor and critics separately
self.actor = ActorNetwork(self.state_size, self.action_size,
self.actor_inter_dims,
use_batch_norm=self.use_batch_norm
).to(self.device)
self.actor_target = ActorNetwork(self.state_size, self.action_size,
self.actor_inter_dims,
use_batch_norm=self.use_batch_norm
).to(self.device)
self.actor_target = utils.copy_weights(self.actor, self.actor_target)
self.critic = CriticNetwork(self.state_size, self.action_size,
self.critic_inter_dims
).to(self.device)
self.critic_target = CriticNetwork(self.state_size, self.action_size,
self.critic_inter_dims
).to(self.device)
self.critic_target = utils.copy_weights(self.critic, self.critic_target)
# initializer optimizers
self.actor_optimizer = optim.Adam(self.actor.parameters(),
lr=self.actor_alpha)
self.critic_optimizer = optim.Adam(self.critic.parameters(),
lr=self.critic_alpha,
eps=1e-4)
# initialize the replay buffer
self.memory = ReplayBuffer(self.buffer_size, self.batch_size,
seed=self.seed)
def step(self):
"""
Update state of the agent and take a step through the learning process
to reflect experiences have been acquired and/or learned from.
"""
# update actor target network
self.actor_target = utils.copy_weights(self.actor, self.actor_target,
self.tau)
# update critic target network
self.critic_target = utils.copy_weights(self.critic,
self.critic_target, self.tau)
def step(self):
"""
Update state of the agent and take a step through the learning process
to reflect experiences have been acquired and/or learned from.
"""
# update actor target network
self.prev_policy = utils.copy_weights(self.policy, self.prev_policy)
# decay epsilon for random noise
self.epsilon = np.max([self.epsilon * self.epsilon_decay,
self.epsilon_min])
# update action variance to choose more selectively
self.action_variances = self.set_action_variances(self.variance_decay)
self.t = -1