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).cpu().to(self.device)
policy_distribution = create_actor_distribution(
self.action_types, policy_output, self.action_size)
actions_tensor = torch.from_numpy(np.array(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().numpy()
if self.action_types == "CONTINUOUS":
action += self.noise.sample()
return action
def produce_action_and_action_info(self, state):
"""Given the state, produces an action, the probability of the action, the log probability of the action, and
the argmax action"""
action_probabilities = self.actor_local(state)
max_probability_action = torch.argmax(action_probabilities).unsqueeze(
0)
action_distribution = create_actor_distribution(
self.action_types, action_probabilities, self.action_size)
action = action_distribution.sample().cpu()
log_action_probabilities = torch.log(action_probabilities)
return action, (action_probabilities,
log_action_probabilities), max_probability_action