def pick_action_and_get_critic_values(self,
policy,
state,
epsilon_exploration=None):
"""Picks an action using the policy"""
# state = torch.from_numpy(state).float().unsqueeze(0)
state = torch.from_numpy(state).float()
model_output = policy.forward(state)
actor_output = model_output[:, list(
range(self.action_size)
)] # we only use first set of columns to decide action, last column is state-value
critic_output = model_output[:, -1]
# action_distribution = create_actor_distribution(self.action_types, actor_output, self.action_size)
action_distribution = create_actor_distribution(
self.action_types, model_output, self.action_size)
action = action_distribution.sample().cpu().numpy()
if self.action_types == "CONTINUOUS": action += self.noise.sample()
if self.action_types == "DISCRETE":
if random.random() <= epsilon_exploration:
action = random.randint(0, self.action_size - 1)
else:
action = action[0]
action_log_prob = self.calculate_log_action_probability(
action, action_distribution)
return action, action_log_prob, critic_output
def calculate_log_probability_of_actions(self, policy, states, actions):
"""Calculates the log probability of an action occuring given a policy and starting state"""
policy_output = policy.forward(states).to(self.device)
policy_distribution = create_actor_distribution(
self.action_types, policy_output, self.action_size)
actions_tensor = actions
policy_distribution_log_prob = policy_distribution.log_prob(
actions_tensor)
return policy_distribution_log_prob
def pick_action(self, policy, state):
"""Picks an action using the policy"""
state = torch.from_numpy(state).float().unsqueeze(0)
actor_output = policy.forward(state)
action_distribution = create_actor_distribution(
self.action_types, actor_output, self.action_size)
action = action_distribution.sample().cpu().numpy()
if self.action_types == "CONTINUOUS":
action += self.noise.sample()
return action
def pick_action(self, policy, state, epsilon_exploration=None):
"""Picks an action using the policy"""
if self.action_types == "DISCRETE":
if random.random() <= epsilon_exploration:
action = random.randint(0, self.action_size - 1)
return action
state = torch.from_numpy(state).float().unsqueeze(0)
actor_output = policy.forward(state)
if self.action_choice_output_columns is not None:
actor_output = actor_output[:, self.action_choice_output_columns]
action_distribution = create_actor_distribution(
self.action_types, actor_output, self.action_size)
action = action_distribution.sample().cpu()
if self.action_types == "CONTINUOUS":
action += torch.Tensor(self.noise.sample())
else:
action = action.item()
return action