def __init__(self, task, gym=False):
self.task = task
if gym:
self.state_size = np.prod(task.observation_space.shape)
self.action_size = np.prod(task.action_space.shape)
self.action_low = task.action_space.low
self.action_high = task.action_space.high
else:
self.state_size = task.state_size
self.action_size = task.action_size
self.action_low = task.action_low
self.action_high = task.action_high
self.alpha = task.alpha
self.beta = task.beta
# Actor (Policy) Model
self.actor_local = Actor(self.state_size, self.action_size, self.action_low, self.action_high, self.alpha)
self.actor_target = Actor(self.state_size, self.action_size, self.action_low, self.action_high, self.alpha)
# Critic (Value) Model
self.critic_local = Critic(self.state_size, self.action_size, self.beta)
self.critic_target = Critic(self.state_size, self.action_size, self.beta)
# Initialize target model parameters with local model parameters
self.critic_target.model.set_weights(self.critic_local.model.get_weights())
self.actor_target.model.set_weights(self.actor_local.model.get_weights())
# Noise process
self.exploration_mu = 0 # changed from 0
self.exploration_theta = 0.15 # changed from 0.15
self.exploration_sigma = 3.2 # changed from 0.2
self.noise = OUNoise(self.action_size, self.exploration_mu, self.exploration_theta, self.exploration_sigma)
# Replay memory
self.buffer_size = 1000 # was 100000 originally
self.batch_size = 64
self.memory = ReplayBuffer(self.buffer_size, self.batch_size)
# Algorithm parameters
self.gamma = 0.99 # discount factor, 0.99
self.tau = .001 # for soft update of target parameters, 0.01 originally
def __init__(self, task):
self.task = task
self.state_size = task.state_size
self.action_size = task.action_size
self.action_low = task.action_low
self.action_high = task.action_high
# Actor (Policy) Model
self.actor_local = Actor(self.state_size, self.action_size,
self.action_low, self.action_high)
self.actor_target = Actor(self.state_size, self.action_size,
self.action_low, self.action_high)
# Critic (Value) Model
self.critic_local = Critic(self.state_size, self.action_size)
self.critic_target = Critic(self.state_size, self.action_size)
# Initialize target model parameters with local model parameters
self.critic_target.model.set_weights(
self.critic_local.model.get_weights())
self.actor_target.model.set_weights(
self.actor_local.model.get_weights())
# Noise process
self.best_score = -np.inf
self.exploration_mu = 0
self.exploration_theta = 0.15
self.exploration_sigma = 0.2
self.noise = OUNoise(self.action_size, self.exploration_mu,
self.exploration_theta, self.exploration_sigma)
# Replay memory
self.buffer_size = 100000
self.batch_size = 64
self.memory = ReplayBuffer(self.buffer_size, self.batch_size)
# Algorithm parameters
self.gamma = 0.5 # discount factor
self.tau = 0.5 # for soft update of target parameters
def __init__(self, task):
self.task = task
self.state_size = task.state_size
self.action_size = task.action_size
self.action_low = task.action_low
self.action_high = task.action_high
# Actor Policy Model
self.actor_local = Actor(self.state_size, self.action_size,
self.action_low, self.action_high)
self.actor_target = Actor(self.state_size, self.action_size,
self.action_low, self.action_high)
# Critic Value Model
self.critic_local = Critic(self.state_size, self.action_size)
self.critic_target = Critic(self.state_size, self.action_size)
self.critic_target.model.set_weights(
self.critic_local.model.get_weights())
self.actor_target.model.set_weights(
self.actor_local.model.get_weights())
# Noise process
self.exploration_mu = 0.0
self.exploration_theta = 0.15 #0.15
self.exploration_sigma = 0.25 #0.2
self.noise = OUNoise(self.action_size, self.exploration_mu,
self.exploration_theta, self.exploration_sigma)
#Replay Memory
self.buffer_size = 100000
self.batch_size = 16
self.memory = ReplayBuffer(self.buffer_size, self.batch_size)
# Algorithm parameters
self.gamma = 0.99
self.tau = 0.01